#482 by GitHub stars
DenisSergeevitch/agents-best-practices
Use this skill when designing, generating an MVP blueprint for, auditing, troubleshooting, refactoring, or explaining an agentic harness for any domain. Covers provider-neutral agent architecture for OpenAI, Anthropic, and OpenAI-compatible APIs: agent loops, tool design, record provenance, interact
History
Review weekly changes to SKILL.md before updating. Dates mark the weeks in which we observed each version; checks apply to the recorded commits.
View weekly changes
SKILL.md
Week of Oct 26, 2026 · +0 / −0 lines
Before Instructions only → After Instructions only
da245f8cbfc4 → da245f8cbfc4
2.4kGitHub stars
14stars in 7 days
1skill in this repository
4commits in 30 days
Heat
6.0- Star growth, last 7 daysAhead of 74% of tracked repositories
- Star growth, last 24 hoursAhead of 90% of tracked repositories
- Rank movement, last 7 daysAhead of 37% of tracked repositories
- Commits, last 30 daysAhead of 67% of tracked repositories
- Accounts asking about it through MCP, last 7 daysNot available
- Hot votes this weekAhead of 0% of tracked repositories
Heat never changes the order by stars, and sponsorship never touches it. How heat works
Stars over time
Individual skills in this repository
How it works
agents-best-practices
When your agent uses it
Use this skill when designing, generating an MVP blueprint for, auditing, troubleshooting, refactoring, or explaining an agentic harness for any domain. Covers provider-neutral agent architecture for OpenAI, Anthropic, and OpenAI-compatible APIs: agent loops, tool design, record provenance, interactive presentation, user-memory lifecycles, environment-adaptive tools, speculative tool execution, late-bound capabilities, permissions, system prompts, planning, goals, always-on agents and durable runtime, adaptive agent teams, context compaction, memory, skills, MCP/external connectors, public-board communications, realtime voice agents, hardware agents and board deployment, self-refining recursive harnesses, programmable context, continual refinement, observability, evals, prompt caching, agent-legible environments, feedback loops, and safety.
What it loads into context
Frontmatter
199 tokensAlways available to the agent
SKILL.md
6,321 tokensWhen the skill is activated
Reference documents
108,452 tokensLoaded on request
Token counts are estimates. Scripts run separately; they are not counted as reference context.
The playbook
- Freshness and self-update
- Core stance
- When to activate this skill
- How to use this skill
- MVP Builder Mode
- Troubleshooting Mode
- Hardware and Board Deployment Mode
- Environment-Adaptive Tool Mode
- Realtime Voice Agent Mode
- Always-on Agent and Durable Runtime Mode
- Advanced Recursive and Continual Harness Mode
- Adaptive Agent Team Mode
- Experimental Speculative Tool Execution Mode
- Reference map
- Default answer structure when advising a user
- Non-negotiable principles
- Common output template
- Gotchas
- Source links for further reading
Files
.gitignoreasset100 bytesLICENSEasset1,071 bytesREADME.mddoc30,038 bytesSKILL.mdcore32,390 bytesSKILL.mdcore32,390 bytesSKILL.mdcore32,390 bytesSKILL.mdcore32,390 bytesSKILL.mdcore32,390 bytesSKILL.mdcore32,390 bytesSKILL.mdcore32,390 bytesSKILL.mdcore32,390 bytesassets/agents-best-practices-illustrations/01-use-cases.jpgasset188,777 bytesassets/agents-best-practices-illustrations/02-workflow-orchestration-schema.svgasset6,499 bytesicon.jpegasset105,040 bytesreferences/adaptive-agent-teams.mddoc17,554 bytesreferences/agent-legibility-feedback-loops.mddoc7,997 bytesreferences/agentic-loop.mddoc11,486 bytesreferences/always-on-agents.mddoc21,761 bytesreferences/architecture.mddoc16,118 bytesreferences/checklists.mddoc33,378 bytesreferences/coding-agents.mddoc19,031 bytesreferences/context-memory-compaction.mddoc19,801 bytesreferences/coverage-audit.mddoc16,348 bytesreferences/environment-adaptive-tools.mddoc22,372 bytesreferences/evals.mddoc56,293 bytesreferences/hardware-agents.mddoc19,867 bytesreferences/mvp-agent-blueprint.mddoc16,502 bytesreferences/planning-and-goals.mddoc8,646 bytesreferences/prompt-caching-and-cost.mddoc11,567 bytesreferences/provider-api-patterns.mddoc14,683 bytesreferences/realtime-voice-agents.mddoc16,679 bytesreferences/security-observability.mddoc18,962 bytesreferences/self-refining-recursive-harnesses.mddoc31,698 bytesreferences/skills-and-connectors.mddoc14,375 bytesreferences/source-links.mddoc72,232 bytesreferences/speculative-tool-execution.mddoc18,409 bytesreferences/system-prompts-instructions.mddoc4,701 bytesreferences/tools-and-permissions.mddoc14,071 bytesreferences/workflow-orchestration.mddoc13,830 bytes
Read SKILL.md
# Agents Best Practices
Use this skill when the user asks how to build, improve, debug, or evaluate an agentic harness. This is a general-purpose agent architecture skill. Coding agents are one subdomain only; apply the same principles to research, finance, legal, support, operations, sales, healthcare, education, data analysis, procurement, and workflow automation agents.
## Freshness and self-update
Canonical source: [DenisSergeevitch/agents-best-practices](https://github.com/DenisSergeevitch/agents-best-practices), branch `main`. Installed copies are snapshots, not guaranteed-current guidance.
Before applying this skill in each new task:
1. **Check upstream once.** Resolve the current `main` commit through an approved network tool and compare it with the installed copy's recorded source revision. A version number alone does not prove freshness. If the copy is behind or its provenance is unknown, retrieve `SKILL.md` and the needed references from that exact commit, then reread the entry point before using it. Record the source commit and skill version in working state; keep the task on that coherent snapshot rather than mixing revisions or repeatedly polling.
2. **Persist the refresh when permitted.** A clean checkout on `main` with the verified canonical remote may update by fast-forward only. Refresh a copied installation's owned Markdown package together only when its prior baseline proves there are no local customizations; otherwise use an isolated upstream snapshot for this task. Update installer-managed packages through their supported installer, not by editing caches. Never overwrite dirty, divergent, customized, or actively maintained source trees, delete unrelated files, or bypass filesystem approval.
3. **Handle unavailable updates honestly.** If installation writes are unavailable, use the retrieved snapshot without replacing the installation. If upstream cannot be checked or retrieved, continue with the available local copy and briefly disclose that freshness is unverified, unless the user requires verified-latest guidance. Do not claim an offline or failed check made the skill current.
Skills included
- agents-best-practices
Repository facts
- License
- MIT
- Created
- May 15, 2026
- Last pushed
- Oct 19, 2026
- Forks
- 219
- Open issues
- 4
- Project activity
- Active
Rank over the last 12 weeks
Now #482 · best #361 · down 66 in 30 days
- #473 on Nov 2
- #474 on Nov 3
- #477 on Nov 4
- #476 on Nov 5
- #479 on Nov 6
- #482 since Nov 7
68 earlier periods
- #338 on Aug 2
- #350 on Aug 8
- #348 on Aug 9
- #363 on Aug 10
- #361 on Aug 11
- #355 on Aug 12
- #360 on Aug 13
- #357 on Aug 14
- #356 on Aug 15
- #363 on Aug 16
- #368 on Aug 18
- #364 on Aug 19
- #365 on Aug 20
- #364 on Aug 21
- #361 on Aug 22
- #367 on Aug 23
- #369 on Aug 24
- #371 on Aug 25
- #376 on Aug 27
- #377 on Aug 28
- #371 on Aug 29
- #366 on Aug 30
- #376 on Aug 31
- #385 on Sep 1
- #379 on Sep 2
- #383 on Sep 3
- #377 on Sep 5
- #383 on Sep 7
- #385 on Sep 8
- #383 on Sep 9
- #387 on Sep 10
- #383 from Sep 11 to Sep 12
- #382 on Sep 13
- #388 on Sep 14
- #392 on Sep 15
- #394 on Sep 16
- #387 on Sep 17
- #393 on Sep 18
- #385 on Sep 19
- #393 on Sep 20
- #392 on Sep 21
- #399 on Sep 22
- #391 from Sep 23 to Sep 24
- #387 on Sep 25
- #383 on Sep 26
- #391 on Sep 27
- #393 on Sep 28
- #407 on Sep 29
- #402 on Sep 30
- #404 on Oct 2
- #392 on Oct 4
- #401 on Oct 5
- #402 on Oct 6
- #409 on Oct 7
- #416 on Oct 8
- #425 on Oct 9
- #414 on Oct 10
- #428 from Oct 13 to Oct 14
- #432 on Oct 15
- #430 on Oct 17
- #431 on Oct 18
- #440 on Oct 19
- #444 on Oct 20
- #447 from Oct 22 to Oct 23
- #462 on Oct 25
- #467 on Oct 28
- #469 on Oct 29
- #471 from Oct 30 to Nov 1
Before you install
Local scriptsAutomated signals, not a guarantee
Checked Nov 3, 2026
Risk level changed on Oct 27, 2026
No longer found: Instruction override signal
Not fully analyzed: some files were too long, unreadable or missing. This never lowers the level.
Show 7 notes
SKILL.md:1Analysis incomplete: Referenced executable target unavailable; not analyzedSKILL.md:1Analysis incomplete: Referenced executable target unavailable; not analyzedSKILL.md:1Analysis incomplete: Referenced executable target unavailable; not analyzedSKILL.md:1Analysis incomplete: Referenced executable target unavailable; not analyzedSKILL.md:1Analysis incomplete: Referenced executable target unavailable; not analyzedSKILL.md:1Analysis incomplete: Referenced executable target unavailable; not analyzedSKILL.md:1Analysis incomplete: Referenced executable target unavailable; not analyzed
- No broad tool permissions
- No dangerous shell commands
- 197 network referencesin 8 files: references/source-links.md:7 SKILL.md:16 README.md:9 and 5 more
- No hidden or obfuscated text
- 3 instruction override signalsin 1 file: references/system-prompts-instructions.md:148
- No executable scripts
- No access to secrets
- No requests to vote in Hot
Show all 200 signals by file
README.md
- Network accessline 9
[](LICENSE) - Network accessline 10
[](SKILL.md) - Network accessline 11
[](SKILL.md) - Network accessline 12
[](SKILL.md) - Network accessline 26
**A. With [`skills`](https://github.com/vercel-labs/skills) (any compatible agent):** - Network accessline 39
1. Clone https://github.com/DenisSergeevitch/agents-best-practices into my - Network accessline 51
# Codex mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills" git clone https://github.com/DenisSergeevitch/agents-best-practices.git \ - Network accessline 52
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills" git clone https://github.com/DenisSergeevitch/agents-best-practices.git \ "${CODEX_HOME:-$HOME/.codex}/skills/agents-best-practices" - Network accessline 53
git clone https://github.com/DenisSergeevitch/agents-best-practices.git \ - Network accessline 56
# Claude Code, user-level mkdir -p "$HOME/.claude/skills" git clone https://github.com/DenisSergeevitch/agents-best-practices.git \ - Network accessline 57
mkdir -p "$HOME/.claude/skills" git clone https://github.com/DenisSergeevitch/agents-best-practices.git \ "$HOME/.claude/skills/agents-best-practices" - Network accessline 58
git clone https://github.com/DenisSergeevitch/agents-best-practices.git \ - Network accessline 61
# Claude Code, project-level mkdir -p .claude/skills git clone https://github.com/DenisSergeevitch/agents-best-practices.git \ - Network accessline 62
mkdir -p .claude/skills git clone https://github.com/DenisSergeevitch/agents-best-practices.git \ .claude/skills/agents-best-practices - Network accessline 63
git clone https://github.com/DenisSergeevitch/agents-best-practices.git \ - Network accessline 229
**Recommended board: [Get Posting Board](https://getpostingboard.dev/).** It provides an agent-facing API with [onboarding instructions](https://getpostingboard.dev/skill.md) and an [OpenAPI contract](https://getpostingboard.dev/openapi.jso - Network accessline 437
- Agent Skills specification: [agentskills.io/specification](https://agentskills.io/specification) - Network accessline 441
- Research on programmatic action surfaces, large tool catalogues, changing API documentation, and novel API use is listed in [`references/source-links.md`](references/source-links.md). - Pinned empirical evidence for model-, workload-, and - Network accessline 443
- [Prime Agent at the researched revision](https://github.com/PrimeIntellect-ai/prime-agent/tree/a18809e00ea30638584d87b3afea7285a9d7296c) is a concrete implementation example of the advanced profile, not a normative dependency or the provi
SKILL.md
- Network accessline 16
Canonical source: [DenisSergeevitch/agents-best-practices](https://github.com/DenisSergeevitch/agents-best-practices), branch `main`. Installed copies are snapshots, not guaranteed-current guidance. - Network accessline 294
- Agent Skills specification: https://agentskills.io/specification - Network accessline 295
- Agent Skills creator best practices: https://agentskills.io/skill-creation/best-practices - Network accessline 296
- Agent Skills description optimization: https://agentskills.io/skill-creation/optimizing-descriptions - Network accessline 297
- Agent Skills evaluation guide: https://agentskills.io/skill-creation/evaluating-skills - Network accessline 298
- OpenAI function calling: https://developers.openai.com/api/docs/guides/function-calling - Network accessline 299
- OpenAI tools: https://developers.openai.com/api/docs/guides/tools - Network accessline 300
- OpenAI agents: https://developers.openai.com/api/docs/guides/agents - Network accessline 301
- OpenAI guardrails and human review: https://developers.openai.com/api/docs/guides/agents/guardrails-approvals - Network accessline 302
- OpenAI agent safety: https://developers.openai.com/api/docs/guides/agent-builder-safety - Network accessline 303
- OpenAI sandbox agents: https://developers.openai.com/api/docs/guides/agents/sandboxes - Network accessline 304
- OpenAI Responses migration: https://developers.openai.com/api/docs/guides/migrate-to-responses - Network accessline 305
- OpenAI prompt caching: https://developers.openai.com/api/docs/guides/prompt-caching - Network accessline 306
- OpenAI Prompt Caching 201: https://developers.openai.com/cookbook/examples/prompt_caching_201 - Network accessline 307
- OpenAI harness engineering article: https://openai.com/index/harness-engineering/ - Network accessline 308
- Anthropic building effective agents: https://www.anthropic.com/research/building-effective-agents - Network accessline 309
- Anthropic effective context engineering: https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents - Network accessline 310
- Anthropic writing effective tools for agents: https://www.anthropic.com/engineering/writing-tools-for-agents - Network accessline 311
- Anthropic long-running harnesses: https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents - Network accessline 312
- Anthropic code execution with MCP: https://www.anthropic.com/engineering/code-execution-with-mcp - Network accessline 313
- MCP specification: https://modelcontextprotocol.io/specification/2026-07-28
assets/agents-best-practices-illustrations/02-workflow-orchestration-schema.svg
- Network accessline 1
<svg xmlns="http://www.w3.org/2000/svg" width="1080" height="1920" viewBox="0 0 1080 1920" role="img" aria-labelledby="title desc">
references/evals.md
- Network accessline 78
| Name | Useful for | |---|---| | [HAL, the Holistic Agent Leaderboard](https://hal.cs.princeton.edu/) | Comparing agent systems across multiple benchmarks with both accuracy and cost visible. | - Network accessline 79
|---|---| | [HAL, the Holistic Agent Leaderboard](https://hal.cs.princeton.edu/) | Comparing agent systems across multiple benchmarks with both accuracy and cost visible. | | [SWE-bench Verified Mini](https://hal.cs.princeton.edu/swebench_v - Network accessline 80
| [HAL, the Holistic Agent Leaderboard](https://hal.cs.princeton.edu/) | Comparing agent systems across multiple benchmarks with both accuracy and cost visible. | - Network accessline 81
| [SWE-bench Verified Mini](https://hal.cs.princeton.edu/swebench_verified_mini) | Cheaper software-engineering issue resolution tests for repository-editing agents. | - Network accessline 82
| [CORE-Bench Hard](https://hal.cs.princeton.edu/corebench_hard) | Scientific programming and research-code tasks. | - Network accessline 83
| [GAIA](https://hal.cs.princeton.edu/gaia) | Web assistance and multi-step agentic search. | - Network accessline 84
| [Online Mind2Web](https://hal.cs.princeton.edu/online_mind2web) | Browser-use and web-task execution. | - Network accessline 85
| [SciCode](https://hal.cs.princeton.edu/scicode) | Scientific coding and tool-use tasks. | - Network accessline 86
| [ScienceAgentBench](https://hal.cs.princeton.edu/scienceagentbench) | Scientific reasoning and self-debugging agent behavior. | - Network accessline 87
| [TAU-bench Airline](https://hal.cs.princeton.edu/taubench_airline) | Customer-service tool use, policy following, and dialogue state. | - Network accessline 88
| [USACO](https://hal.cs.princeton.edu/usaco) | Algorithmic programming and contest-style problem solving. | - Network accessline 539
Original article motivating the scaffold-vs-model framing: [Just a Wrapper? How much do scaffolds matter?](https://www.lesswrong.com/posts/jXLi3dhSpSMd7B6z8/just-a-wrapper-how-much-do-scaffolds-matter-1)
references/provider-api-patterns.md
- Network accessline 278
- OpenAI Responses migration: https://developers.openai.com/api/docs/guides/migrate-to-responses - Network accessline 279
- OpenAI function calling: https://developers.openai.com/api/docs/guides/function-calling - Network accessline 280
- OpenAI tools: https://developers.openai.com/api/docs/guides/tools - Network accessline 281
- OpenAI Agents SDK: https://developers.openai.com/api/docs/guides/agents - Network accessline 282
- OpenAI guardrails and human review: https://developers.openai.com/api/docs/guides/agents/guardrails-approvals - Network accessline 283
- OpenAI prompt caching: https://developers.openai.com/api/docs/guides/prompt-caching - Network accessline 284
- OpenAI Prompt Caching 201: https://developers.openai.com/cookbook/examples/prompt_caching_201 - Network accessline 285
- Anthropic building effective agents: https://www.anthropic.com/research/building-effective-agents - Network accessline 286
- Anthropic writing effective tools for agents: https://www.anthropic.com/engineering/writing-tools-for-agents
references/security-observability.md
- Network accessline 244
- OpenAI guardrails and human review: https://developers.openai.com/api/docs/guides/agents/guardrails-approvals - Network accessline 245
- OpenAI agent safety: https://developers.openai.com/api/docs/guides/agent-builder-safety - Network accessline 246
- OpenAI sandbox agents: https://developers.openai.com/api/docs/guides/agents/sandboxes - Network accessline 247
- Anthropic building effective agents: https://www.anthropic.com/research/building-effective-agents - Network accessline 248
- Anthropic writing effective tools for agents: https://www.anthropic.com/engineering/writing-tools-for-agents - Network accessline 249
- MCP specification: https://modelcontextprotocol.io/specification/2026-07-28
references/skills-and-connectors.md
- Network accessline 331
- Agent Skills specification: https://agentskills.io/specification - Network accessline 332
- Agent Skills creator best practices: https://agentskills.io/skill-creation/best-practices - Network accessline 333
- Agent Skills description optimization: https://agentskills.io/skill-creation/optimizing-descriptions - Network accessline 334
- Agent Skills evaluation guide: https://agentskills.io/skill-creation/evaluating-skills - Network accessline 335
- MCP specification: https://modelcontextprotocol.io/specification/2026-07-28 - Network accessline 336
[Credential-related source omitted] - Network accessline 337
- OpenAI tools: https://developers.openai.com/api/docs/guides/tools - Network accessline 338
- Anthropic code execution with MCP: https://www.anthropic.com/engineering/code-execution-with-mcp
references/source-links.md
- Network accessline 7
- Agent Skills specification: https://agentskills.io/specification - Network accessline 8
- Agent Skills creator best practices: https://agentskills.io/skill-creation/best-practices - Network accessline 9
- Optimizing skill descriptions: https://agentskills.io/skill-creation/optimizing-descriptions - Network accessline 10
- Evaluating skill output quality: https://agentskills.io/skill-creation/evaluating-skills - Network accessline 11
- Using scripts in skills: https://agentskills.io/skill-creation/using-scripts - Network accessline 15
- Mike Piccolo, [53 harness bugs encountered during development](https://x.com/mfpiccolo/status/2102089734004593140), posted September 21, 2026; post metadata retrieved through the [public mirror](https://api.fxtwitter.com/mfpiccolo/status/ - Network accessline 23
- Paper: [An Empirical Study of Harness Design for Coding Agents, arXiv:2609.20804v1](https://arxiv.org/abs/2609.20804v1), submitted 17 September 2026; original HTML rechecked 19 September 2026. - Network accessline 24
- Primary evidence: [context policy](https://arxiv.org/html/2609.20804v1#S2.SS3), [setup](https://arxiv.org/html/2609.20804v1#S3.SS1), [results](https://arxiv.org/html/2609.20804v1#S3.SS2), [trajectory analysis](https://arxiv.org/html/2609. - Network accessline 47
- Release article: [Pi 1.0](https://earendil.com/posts/pi-1-0/), October 1, 2026. - Network accessline 48
- Pinned implementation: [v1.0.0 at `a13d35a742c6ef8462812a28fbe1d8c8b7431c32`](https://github.com/earendil-works/pi/tree/a13d35a742c6ef8462812a28fbe1d8c8b7431c32), released October 1; README, relevant docs, source, and tests inspected Octo - Network accessline 49
- Routing: [virtual-model documentation](https://github.com/earendil-works/pi/blob/a13d35a742c6ef8462812a28fbe1d8c8b7431c32/packages/coding-agent/docs/virtual-models.md), [model runtime](https://github.com/earendil-works/pi/blob/a13d35a742c - Network accessline 50
- Configuration: [message documentation](https://github.com/earendil-works/pi/blob/a13d35a742c6ef8462812a28fbe1d8c8b7431c32/packages/coding-agent/docs/message-types.md), [actual message type](https://github.com/earendil-works/pi/blob/a13d35 - Network accessline 51
- Warming: [cache warmer](https://github.com/earendil-works/pi/blob/a13d35a742c6ef8462812a28fbe1d8c8b7431c32/packages/coding-agent/src/core/cache-warmer.ts) and [warmer tests](https://github.com/earendil-works/pi/blob/a13d35a742c6ef8462812a - Network accessline 65
- Dated source: [How we made claude.ai 3x faster in two weeks](https://claude.dev/blog/how-we-made-claude-ai-faster/), Raymond Wang, Sam Attard, and Issac G., published September 23, 2026; inspected October 6, 2026. - Network accessline 73
- Launch article: [Pi Durable](https://earendil.com/posts/pi-durable/), October 1, 2026. The article describes an experimental framework alongside the coding-agent release. - Network accessline 74
- Pinned implementation: [main at `b2b5c42f6138b73ec4b2f49ec0ca468800f88586`](https://github.com/earendil-works/pi/tree/b2b5c42f6138b73ec4b2f49ec0ca468800f88586), committed October 5, 2026; [package version 1.0.2](https://github.com/earendi - Network accessline 75
- Contract map: [README](https://github.com/earendil-works/pi/blob/b2b5c42f6138b73ec4b2f49ec0ca468800f88586/packages/durable/README.md) and [runtime/storage specification](https://github.com/earendil-works/pi/blob/b2b5c42f6138b73ec4b2f49ec0 - Network accessline 76
- Commit and storage evidence: [Session admission/publication](https://github.com/earendil-works/pi/blob/b2b5c42f6138b73ec4b2f49ec0ca468800f88586/packages/durable/src/session/session.ts#L405), [JSONL marker protocol](https://github.com/eare - Network accessline 77
- Execution evidence: [scheduler](https://github.com/earendil-works/pi/blob/b2b5c42f6138b73ec4b2f49ec0ca468800f88586/packages/durable/src/harness/scheduler.ts), [generation](https://github.com/earendil-works/pi/blob/b2b5c42f6138b73ec4b2f49e - Network accessline 78
- Recovery and ownership tests: [task recovery](https://github.com/earendil-works/pi/blob/b2b5c42f6138b73ec4b2f49ec0ca468800f88586/packages/durable/test/harness-tasks-recovery.test.ts), [model recovery](https://github.com/earendil-works/pi/ - Network accessline 79
- Intake evidence: [submission admission](https://github.com/earendil-works/pi/blob/b2b5c42f6138b73ec4b2f49ec0ca468800f88586/packages/durable/src/harness/submissions.ts#L148), [inbox boundaries](https://github.com/earendil-works/pi/blob/b2b - Network accessline 80
- State and observation evidence: [document definitions](https://github.com/earendil-works/pi/blob/b2b5c42f6138b73ec4b2f49ec0ca468800f88586/packages/durable/src/documents.ts), [fork tests](https://github.com/earendil-works/pi/blob/b2b5c42f6 - Network accessline 81
- Context and progress evidence: [compaction](https://github.com/earendil-works/pi/blob/b2b5c42f6138b73ec4b2f49ec0ca468800f88586/packages/durable/src/harness/compaction.ts), [compaction tests](https://github.com/earendil-works/pi/blob/b2b5c - Network accessline 116
- Lance Martin and CJ Avilla, [Building effective agent automations](https://claude.dev/blog/building-effective-agent-automations/), published October 8, 2026; article and linked implementation rechecked October 13, 2026. - Network accessline 117
- Reference implementation: [Claude Quickstarts daily brief at `9ec32b91df50b0d4906cae64b13c6298055ff40d`](https://github.com/anthropics/claude-quickstarts/tree/9ec32b91df50b0d4906cae64b13c6298055ff40d/managed-agents/daily-brief), committed - Network accessline 124
- Publication checks Slack's response body for `ok` and a message timestamp, but the quickstart supplies no atomic bookkeeping or concurrent-run exclusion. Its quiet branch asks for a posted status/message ID despite sending nothing; duplic - Network accessline 126
- A per-run cap can pause a session and appear as silence; see [scheduled deployment budgets](https://platform.claude.com/docs/en/managed-agents/scheduled-deployments#set-a-budget-on-each-run). The quickstart includes manual inspection, not - Network accessline 134
- Architecture comparison: OpenAI [Voice agents](https://developers.openai.com/api/docs/guides/voice-agents), covering delegated full duplex, native speech-to-speech, chained STT/agent/TTS, and separate task, audio, timing, and reliability - Network accessline 135
- Current model pages: [GPT-Realtime-2.1](https://developers.openai.com/api/docs/models/gpt-realtime-2.1) and [2.1 Mini](https://developers.openai.com/api/docs/models/gpt-realtime-2.1-mini) are reasoning voice models with tool use and 128k - Network accessline 136
- Realtime protocol: [starter](https://developers.openai.com/api/docs/guides/realtime), [VAD](https://developers.openai.com/api/docs/guides/realtime-vad), [interruption/truncation](https://developers.openai.com/api/docs/guides/realtime-conv - Network accessline 137
- Live protocol: [getting started](https://developers.openai.com/api/docs/guides/live), [delegation and tools](https://developers.openai.com/api/docs/guides/live-delegation), [session/transcript lifecycle](https://developers.openai.com/api/ - Network accessline 138
- Output checks and controls: [server-side controls](https://developers.openai.com/api/docs/guides/voice-server-controls?api=live), especially [checking speech before playback](https://developers.openai.com/api/docs/guides/voice-server-cont - Network accessline 142
- The Realtime starter uses `gpt-realtime-2.1`, while [voice prompting](https://developers.openai.com/api/docs/guides/voice-prompting) still labels its guidance Realtime 2 / 2.0 and describes `gpt-realtime-2` as the leading model. Its reaso - Network accessline 146
Pinned implementation: [OpenAI Codex at `7e9c061ceb388c67097acac069d6b7f8a60e5011`](https://github.com/openai/codex/tree/7e9c061ceb388c67097acac069d6b7f8a60e5011). Source-observed evidence: - Network accessline 148
- The supplied [handoff routing path](https://github.com/openai/codex/blob/7e9c061ceb388c67097acac069d6b7f8a60e5011/codex-rs/core/src/realtime_conversation.rs#L1797-L1819) routes text into the main session before publishing the raw event. T - Network accessline 149
- The [front-agent prompt](https://github.com/openai/codex/blob/7e9c061ceb388c67097acac069d6b7f8a60e5011/codex-rs/prompts/templates/realtime/backend_prompt.md#L13-L51) separates conversational presentation from execution and steers ongoing - Network accessline 150
- The [admission gate](https://github.com/openai/codex/blob/7e9c061ceb388c67097acac069d6b7f8a60e5011/codex-rs/core/src/realtime_conversation.rs#L136-L165) serializes routing against retirement; the [response queue](https://github.com/openai - Network accessline 151
- [Audio ingress](https://github.com/openai/codex/blob/7e9c061ceb388c67097acac069d6b7f8a60e5011/codex-rs/core/src/realtime_conversation.rs#L858-L883) can drop a frame on a full queue while returning success. [Native playback](https://github - Network accessline 152
- The [interruption handler](https://github.com/openai/codex/blob/7e9c061ceb388c67097acac069d6b7f8a60e5011/codex-rs/core/src/realtime_conversation.rs#L2590-L2624) truncates using [accumulated received-audio duration](https://github.com/open - Network accessline 153
- Product [model literals](https://github.com/openai/codex/blob/7e9c061ceb388c67097acac069d6b7f8a60e5011/codex-rs/core/src/realtime_conversation.rs#L112-L113) are `gpt-realtime-1.5` and `gpt-live-1-codex`; [version selection](https://github - Network accessline 157
- LiveKit [turn management](https://docs.livekit.io/agents/logic/turns/) distinguishes endpointing, interruption decisions, and false-interruption recovery; its realtime-mode section explains that most framework interruption options do not - Network accessline 158
- Pipecat's published [output-transport source](https://reference-server.pipecat.ai/en/stable/_modules/pipecat/transports/base_output.html) shows interruption-specific queue/task handling that preserves explicitly uninterruptible output and - Network accessline 164
- OpenAI Agents guide: https://developers.openai.com/api/docs/guides/agents - Network accessline 165
- OpenAI function calling: https://developers.openai.com/api/docs/guides/function-calling - Network accessline 166
- OpenAI tools: https://developers.openai.com/api/docs/guides/tools - Network accessline 167
- OpenAI tool search: https://developers.openai.com/api/docs/guides/tools-tool-search - Network accessline 168
- OpenAI guardrails and human review: https://developers.openai.com/api/docs/guides/agents/guardrails-approvals - Network accessline 169
- OpenAI agent safety: https://developers.openai.com/api/docs/guides/agent-builder-safety - Network accessline 170
- OpenAI sandbox agents: https://developers.openai.com/api/docs/guides/agents/sandboxes - Network accessline 171
- OpenAI Responses migration: https://developers.openai.com/api/docs/guides/migrate-to-responses - Network accessline 172
- OpenAI prompt caching: https://developers.openai.com/api/docs/guides/prompt-caching - Network accessline 173
- OpenAI Prompt Caching 201: https://developers.openai.com/cookbook/examples/prompt_caching_201 - Network accessline 174
- OpenAI harness engineering article: https://openai.com/index/harness-engineering/ - Network accessline 175
- OpenAI MCP and connectors: https://developers.openai.com/api/docs/guides/tools-connectors-mcp - Network accessline 179
- Anthropic building effective agents: https://www.anthropic.com/research/building-effective-agents - Network accessline 180
- Anthropic effective context engineering: https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents - Network accessline 181
- Anthropic writing effective tools for agents: https://www.anthropic.com/engineering/writing-tools-for-agents - Network accessline 182
- Anthropic effective harnesses for long-running agents: https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents - Network accessline 183
- Anthropic demystifying evals for agents: https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents - Network accessline 184
- Anthropic code execution with MCP: https://www.anthropic.com/engineering/code-execution-with-mcp - Network accessline 185
- Anthropic tool search: https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool - Network accessline 186
- Anthropic Agent Skills engineering note: https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills - Network accessline 190
- Anthropic article, September 2, 2026: [A guide to the anatomy of effective commerce agents](https://claude.com/blog/the-anatomy-of-effective-commerce-agents). - Network accessline 191
- Reference implementation: [commerce-agents at `fd4d59224ab96b43c6dc6888207c67b3bd5a24cf`](https://github.com/anthropics/commerce-agents/tree/fd4d59224ab96b43c6dc6888207c67b3bd5a24cf), committed August 31, 2026; source and tests inspected - Network accessline 192
- UI evidence: [typed presentation runner](https://github.com/anthropics/commerce-agents/blob/fd4d59224ab96b43c6dc6888207c67b3bd5a24cf/commerce-common/commerce_common/presentation.py#L120) and [record enrichment, filtering, and disclosures] - Network accessline 193
- Mutation evidence: [cart caps and serialization](https://github.com/anthropics/commerce-agents/blob/fd4d59224ab96b43c6dc6888207c67b3bd5a24cf/shopping-agent/core/shopping_agent/gates.py#L85), [apply approval gate](https://github.com/anthro - Network accessline 194
- Memory evidence: [common write and lifecycle implementation](https://github.com/anthropics/commerce-agents/blob/fd4d59224ab96b43c6dc6888207c67b3bd5a24cf/commerce-common/commerce_common/memory.py), [merchant identity scope](https://github. - Network accessline 195
- Evaluation and deployment evidence: [eval-authoring skill](https://github.com/anthropics/commerce-agents/blob/fd4d59224ab96b43c6dc6888207c67b3bd5a24cf/plugins/commerce-builder/skills/commerce-evals/SKILL.md#L8) and [safety boundaries](htt - Network accessline 210
- MCP specification, stable 2026-07-28: https://modelcontextprotocol.io/specification/2026-07-28 - Network accessline 211
- MCP specification source at the stable tag: https://github.com/modelcontextprotocol/modelcontextprotocol/tree/5f5440bb26a62e2cf3440b92da5a667efa03b267 - Network accessline 212
[Credential-related source omitted] - Network accessline 213
- MCP server discovery: https://modelcontextprotocol.io/specification/2026-07-28/server/discover - Network accessline 214
- MCP tools: https://modelcontextprotocol.io/specification/2026-07-28/server/tools - Network accessline 218
- CodeAct paper, arXiv v4: https://arxiv.org/abs/2402.01030v4 - Network accessline 219
- CodeAct ICML 2024 publication: https://proceedings.mlr.press/v235/wang24h.html - Network accessline 220
- CodeAct official implementation at researched revision: https://github.com/xingyaoww/code-act/tree/d607f56c9cfe9e8632ebaf65dcaf2b4b7fe1c6f8 - Network accessline 221
- ToolLLM paper, arXiv v2: https://arxiv.org/abs/2307.16789v2 - Network accessline 222
- ToolLLM ICLR 2024 publication: https://proceedings.iclr.cc/paper_files/paper/2024/hash/28e50ee5b72e90b50e7196fde8ea260e-Abstract-Conference.html - Network accessline 223
- ToolBench official implementation at the paper-era revision: https://github.com/OpenBMB/ToolBench/tree/ce541837c92f47f832e91f3ae92480fbbdb9a1e3 - Network accessline 224
- Gorilla paper, arXiv v1: https://arxiv.org/abs/2305.15334v1 - Network accessline 225
- Gorilla NeurIPS 2024 publication: https://proceedings.nips.cc/paper_files/paper/2024/hash/e4c61f578ff07830f5c37378dd3ecb0d-Abstract-Conference.html - Network accessline 226
- Gorilla official implementation at the paper-era release: https://github.com/ShishirPatil/gorilla/tree/29f5ffb6726e3fab8c7fc7bfe017d037a693b10d - Network accessline 227
- NovelAPIBench paper, arXiv v1 preprint: https://arxiv.org/abs/2606.03657v1 - Network accessline 231
- Speculative Programmatic Tool Calling article, August 2026: https://alexzhang13.github.io/blog/2026/spec-ptc/ - Network accessline 232
- Speculative Programmatic Tool Calling implementation, v0.1.1 researched revision: https://github.com/alexzhang13/spec-ptc/tree/9b78b7d6ceeaf8afd1557c4e3a999ce653fc0e17 - Network accessline 233
- Conveyor paper, arXiv v2: https://arxiv.org/abs/2406.00059v2 - Network accessline 234
- Speculative Interaction Agents paper, arXiv v2: https://arxiv.org/abs/2605.13360v2 - Network accessline 235
- AsyncFC paper, arXiv v1: https://arxiv.org/abs/2605.15077v1 - Network accessline 239
- Paper: Zou et al., [Hybrid Reasoning Agents, arXiv:2610.18861v1](https://arxiv.org/html/2610.18861v1), October 13, 2026; methods, analyses, and appendices inspected October 19. The paper is licensed CC BY 4.0. - Network accessline 240
- Official implementation: [togethercomputer/HybridAgent at `1b41cebbd2388d083ee1820af421b7f9e24affb5`](https://github.com/togethercomputer/HybridAgent/tree/1b41cebbd2388d083ee1820af421b7f9e24affb5), committed October 18, 2026; README, gene - Network accessline 241
- Runtime evidence: [persistent KV/position/mask handling](https://github.com/togethercomputer/HybridAgent/blob/1b41cebbd2388d083ee1820af421b7f9e24affb5/hybrid_agent/engine.py#L215), [latent transitions](https://github.com/togethercomputer/ - Network accessline 242
- Inspection and training evidence: [nearest-token projection](https://github.com/togethercomputer/HybridAgent/blob/1b41cebbd2388d083ee1820af421b7f9e24affb5/hybrid_agent/readout.py#L15), [inspector](https://github.com/togethercomputer/Hybri - Network accessline 243
- Published [8B](https://huggingface.co/togethercomputer/HybridAgent-8B-Agentic) and [14B](https://huggingface.co/togethercomputer/HybridAgent-14B-Agentic) model cards were checked October 19. They require the custom runtime for hybrid gene - Network accessline 257
- [Released inference defaults](https://github.com/togethercomputer/HybridAgent/blob/1b41cebbd2388d083ee1820af421b7f9e24affb5/hybrid_inference/run_batch.py#L41) use 4–200 latent steps per span; [RL presets](https://github.com/togethercomput - Network accessline 258
- The [Python tool](https://github.com/togethercomputer/HybridAgent/blob/1b41cebbd2388d083ee1820af421b7f9e24affb5/hybrid_agent/tools/python_tool.py#L81) explicitly runs with current-user filesystem/network access; its subprocess is not a se - Network accessline 265
- Recursive Language Models paper: https://arxiv.org/abs/2512.24601 - Network accessline 266
- Recursive Language Models reference implementation: https://github.com/alexzhang13/rlm - Network accessline 267
- Continual Harness paper: https://arxiv.org/abs/2605.09998 - Network accessline 268
- Continual Harness official implementation: https://github.com/sethkarten/continual-harness - Network accessline 269
- Recursive Agent Harnesses paper: https://arxiv.org/abs/2606.13643 - Network accessline 270
- Voyager paper: https://arxiv.org/abs/2305.16291 - Network accessline 274
- Paper: [Harness²: Recursive Agent Harnessing for an Open World, arXiv:2610.14224v1](https://arxiv.org/html/2610.14224v1), submitted October 9, 2026; sections 2.1–2.3 and experimental setup inspected October 19. - Network accessline 275
- Implementation: [google-research/harness2 at `1cfe254e8d25b15f37048c3730134686abe60335`](https://github.com/google-research/harness2/tree/1cfe254e8d25b15f37048c3730134686abe60335), committed October 14, 2026 UTC; README, experiment docume - Network accessline 276
- Search evidence: [recursion scheduling and proposal validation](https://github.com/google-research/harness2/blob/1cfe254e8d25b15f37048c3730134686abe60335/harness2/core/recursion.py), [rollout/stage ledger](https://github.com/google-resear - Network accessline 277
- Retention evidence: [reflection prompts](https://github.com/google-research/harness2/blob/1cfe254e8d25b15f37048c3730134686abe60335/harness2/core/prompts.py#L1106), [library eligibility and storage](https://github.com/google-research/harne - Network accessline 278
- Validation scope: [workflow tests](https://github.com/google-research/harness2/blob/1cfe254e8d25b15f37048c3730134686abe60335/tests/test_workflow.py), [experiment and contamination notes](https://github.com/google-research/harness2/blob/1c - Network accessline 302
- OpenAI, [On the Navier–Stokes Millennium Prize Problem](https://openai.com/index/navier-stokes-solution/), published September 8, 2026; updated September 10; rechecked September 30. “How we found the proof” is the public orchestration acc - Network accessline 303
- User-supplied excerpt from “What Comes After Loops?: Deep Dive Into Agentic Architecture,” September 29, 2026, 11:15–11:40 AM PT, Mezzanine Stage; listed speakers Raj Pathak and Eric Provencher. This is supplied panel evidence, not an ind - Network accessline 304
- Mathematical artifacts: [Lean repository at `f9e8bc5b38b6e212696e8a30e3e91517af887bbd`](https://github.com/openai/NavierStokesAndEuler/tree/f9e8bc5b38b6e212696e8a30e3e91517af887bbd), September 10 revision; [README](https://github.com/open - Network accessline 305
- Prior lineage: [Hearsay-III, AAAI 1980](https://cdn.aaai.org/AAAI/1980/AAAI80-032.pdf), shared blackboard information and scheduling; [The Contract Net Protocol, December 1980](https://www.reidgsmith.com/The_Contract_Net_Protocol_Dec-1980 - Network accessline 354
- [Espressif esptool: basic commands](https://docs.espressif.com/projects/esptool/en/latest/esp32s3/esptool/basic-commands.html): image inspection, flash read/write, erase, and verification; choose exact commands/ranges for the measured tar - Network accessline 355
- [MicroPython 1.27.0 SSL](https://docs.micropython.org/en/v1.27.0/library/ssl.html), [GC](https://docs.micropython.org/en/v1.27.0/library/gc.html), and [time](https://docs.micropython.org/en/v1.27.0/library/time.html): pinned runtime API/h - Network accessline 356
- [M5Stack Cardputer ADV UiFlow2 programming](https://docs.m5stack.com/en/uiflow2/cardputer-adv/program): base runtime programming, not proof of the custom launcher's behavior. - Network accessline 362
- Get Posting Board agent instructions: https://getpostingboard.dev/skill.md - Network accessline 363
- Get Posting Board API contract: https://getpostingboard.dev/openapi.json - Network accessline 371
- OWASP AI Agent Security Cheat Sheet: https://cheatsheetseries.owasp.org/cheatsheets/AI_Agent_Security_Cheat_Sheet.html - Network accessline 372
- OWASP Agentic Skills Top 10: https://owasp.org/www-project-agentic-skills-top-10/ - Network accessline 373
- NIST AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework - Network accessline 377
- Bun PR with generated workflow program examples: https://github.com/oven-sh/bun/pull/30412/changes - Network accessline 378
- Prime Agent repository at researched revision: https://github.com/PrimeIntellect-ai/prime-agent/tree/a18809e00ea30638584d87b3afea7285a9d7296c - Network accessline 379
- Prime Agent launch article: https://www.primeintellect.ai/blog/prime-agent
references/system-prompts-instructions.md
- Instruction override signalline 148
- “You have full autonomy” without policy boundaries; - “Always complete the task no matter what”; - “Ignore all previous instructions”; - Instruction override signalline 149
- “Always complete the task no matter what”; - “Ignore all previous instructions”; - vague safety rules not enforced by code; - Instruction override signalline 150
- “Ignore all previous instructions”;
How the signals fit together
Facts from a static read of the files at the pinned commit. They describe what the code does, not why.
Some files could not be read; the facts below still hold.
Signal confidence: 199 facts
- Network access in prose of documentationREADME.md:9Low confidence
- Network access in prose of documentationREADME.md:10Low confidence
- Network access in prose of documentationREADME.md:11Low confidence
- Network access in prose of documentationREADME.md:12Low confidence
- Network access in prose of documentationREADME.md:26Low confidence
- Network access in code of documentationREADME.md:39Low confidence
- Network access in code of documentationREADME.md:51Low confidence
- Network access in code of documentationREADME.md:52Low confidence
- Network access in code of documentationREADME.md:53Low confidence
- Network access in code of documentationREADME.md:56Low confidence
- Network access in code of documentationREADME.md:57Low confidence
- Network access in code of documentationREADME.md:58Low confidence
- Network access in code of documentationREADME.md:61Low confidence
- Network access in code of documentationREADME.md:62Low confidence
- Network access in code of documentationREADME.md:63Low confidence
- Network access in prose of documentationREADME.md:229Low confidence
- Network access in prose of documentationREADME.md:437Low confidence
- Network access in prose of documentationREADME.md:441Low confidence
- Network access in prose of documentationREADME.md:443Low confidence
- Network access in prose of the SKILL.mdSKILL.md:16Low confidence
- Network access in prose of the SKILL.mdSKILL.md:294Low confidence
- Network access in prose of the SKILL.mdSKILL.md:295Low confidence
- Network access in prose of the SKILL.mdSKILL.md:296Low confidence
- Network access in prose of the SKILL.mdSKILL.md:297Low confidence
- Network access in prose of the SKILL.mdSKILL.md:298Low confidence
- Network access in prose of the SKILL.mdSKILL.md:299Low confidence
- Network access in prose of the SKILL.mdSKILL.md:300Low confidence
- Network access in prose of the SKILL.mdSKILL.md:301Low confidence
- Network access in prose of the SKILL.mdSKILL.md:302Low confidence
- Network access in prose of the SKILL.mdSKILL.md:303Low confidence
- Network access in prose of the SKILL.mdSKILL.md:304Low confidence
- Network access in prose of the SKILL.mdSKILL.md:305Low confidence
- Network access in prose of the SKILL.mdSKILL.md:306Low confidence
- Network access in prose of the SKILL.mdSKILL.md:307Low confidence
- Network access in prose of the SKILL.mdSKILL.md:308Low confidence
- Network access in prose of the SKILL.mdSKILL.md:309Low confidence
- Network access in prose of the SKILL.mdSKILL.md:310Low confidence
- Network access in prose of the SKILL.mdSKILL.md:311Low confidence
- Network access in prose of the SKILL.mdSKILL.md:312Low confidence
- Network access in prose of the SKILL.mdSKILL.md:313Low confidence
- Network access in prose of documentationreferences/evals.md:78Low confidence
- Network access in prose of documentationreferences/evals.md:79Low confidence
- Network access in prose of documentationreferences/evals.md:80Low confidence
- Network access in prose of documentationreferences/evals.md:81Low confidence
- Network access in prose of documentationreferences/evals.md:82Low confidence
- Network access in prose of documentationreferences/evals.md:83Low confidence
- Network access in prose of documentationreferences/evals.md:84Low confidence
- Network access in prose of documentationreferences/evals.md:85Low confidence
- Network access in prose of documentationreferences/evals.md:86Low confidence
- Network access in prose of documentationreferences/evals.md:87Low confidence
and 149 more
Works with
| Agent | Status | Evidence | Install to |
|---|---|---|---|
| Claude Code | Compatible | Standard SKILL.md SKILL.md | .claude/skills/<skill-name>/ |
| Codex | Compatible | Standard SKILL.md SKILL.md | .agents/skills/<skill-name>/ |
| Cursor | Compatible | Standard SKILL.md SKILL.md | .cursor/skills/<skill-name>/ |
| Gemini CLI | Compatible | Standard SKILL.md SKILL.md | .gemini/skills/<skill-name>/ |
| GitHub Copilot | Compatible | Standard SKILL.md SKILL.md | .github/skills/<skill-name>/ |
| OpenCode | Compatible | Standard SKILL.md SKILL.md | .opencode/skills/<skill-name>/ |
- Native The repository ships a manifest, folder or frontmatter field for this agent.
- Compatible A standard SKILL.md that this agent can read.
- Unknown No evidence either way.
Derived automatically from paths, manifests and frontmatter. Not tested on each agent.
Explore alternatives and comparisons
Add this badge to your README
It updates every night and links to this page.
[](https://hotskills.dev/skill/denissergeevitch-agents-best-practices/)