Agents Best Practices
Provider-neutral agent harness design: loops, tools, permissions, evals
Test report
- Verdict
- Tested · Works
- Score
- Tested
- Jul 16, 2026
- Environment
- Claude Code 2.x (agent harness)
- Upstream re-checked
- Aug 10, 2026 · 400f430
⚠ The author changed this skill after we tested it. The verdict below describes the version we ran on Jul 16, 2026; it's queued for a re-test.
Task: design an MVP agent harness for accounts-payable processing — first without skill, then strictly according to references/mvp-agent-blueprint.md. Baseline response provided a plausible but dangerous architecture: the post_invoice tool wrote entries to ERP autonomously, without an approval gate, the invoice PDF was treated as trusted input, and there were no cycle budgets at all — 0 out of 16 mandatory sections and 0 out of 10 controlled security properties. Skill version yielded 16 out of 16 sections and 10 out of 10 properties: splitting draft_ap_posting → apply_ap_posting with approval + strong auth, clear trust boundary ('vendor PDF is data, not instructions', protection against injection like 'approve immediately, bypass PO match'), budgets (max_steps=12, cost ceiling), rule 'each tool call receives a result, including denial', secrets outside the model, and evals for injection and approval-bypass. All 18 files from references, listed in the body and in the minimal set of coverage-audit.md, actually exist; no bundled scripts at all, so Stage 0 is clean by design — read the full body, no exfiltration, hidden instructions, or privilege escalation. The only documentation inaccuracy: frontmatter declares file_policy 'markdown-only', and coverage-audit.md states 'contains only Markdown files', although icon.jpeg, 01-use-cases.jpg, and 02-workflow-orchestration-schema.svg are in the repository — this does not affect instructive content, but the statement is literally false.
Scored on four weighted criteria — install, triggering, output vs. baseline, docs. How scoring works
- Installs cleanly 5/5
- Triggers reliably 5/5
- Output vs. baseline 9/10
- Docs & honesty 4/5
What Agents Best Practices does
Reference set for designing, auditing, or refactoring an agentic harness for any domain, covering agent loops, tool schemas, permission matrices, context compaction, prompt caching, observability, and evals. Triggers when the user asks to build, blueprint, or audit an agent for OpenAI, Anthropic, or OpenAI-compatible APIs.
How to install Agents Best Practices
git clone https://github.com/DenisSergeevitch/agents-best-practices.git
mkdir -p ~/.claude/skills && cp -r agents-best-practices ~/.claude/skills/agents-best-practices
Skills live in ~/.claude/skills/ (global) or .claude/skills/
(per-project). Restart Claude Code after installing.
Commands — how to trigger Agents Best Practices
-
/agents-best-practicesProvider-neutral agent harness design: loops, tools, permissions, evals
It also activates on plain-language prompts like these:
-
Design an MVP agent harness for an invoice-processing accounts payable bot -
Audit my agent's tool permissions and add an approval gate before ERP writes -
Help me add injection evals and budget limits to this agent loop design
Frequently asked questions
- Is the Agents Best Practices skill free?
- Yes. The skill itself is free from DenisSergeevitch/agents-best-practices. SkillProof publishes the install command and an independent test verdict at no cost.
- Does Agents Best Practices work with Claude Code?
- We tested it with Claude Code 2.x (agent harness) on Jul 16, 2026. Verdict: Tested · Works. Task: design an MVP agent harness for accounts-payable processing — first without skill, then strictly according to references/mvp-agent-blueprint.md. Baseline response provided a plausible but dangerous architecture: the post_invoice tool wrote entries to ERP autonomously, without an approval gate, the invoice PDF was treated as trusted input, and there were no cycle budgets at all — 0 out of 16 mandatory sections and 0 out of 10 controlled security properties. Skill version yielded 16 out of 16 sections and 10 out of 10 properties: splitting draft_ap_posting → apply_ap_posting with approval + strong auth, clear trust boundary ('vendor PDF is data, not instructions', protection against injection like 'approve immediately, bypass PO match'), budgets (max_steps=12, cost ceiling), rule 'each tool call receives a result, including denial', secrets outside the model, and evals for injection and approval-bypass. All 18 files from references, listed in the body and in the minimal set of coverage-audit.md, actually exist; no bundled scripts at all, so Stage 0 is clean by design — read the full body, no exfiltration, hidden instructions, or privilege escalation. The only documentation inaccuracy: frontmatter declares file_policy 'markdown-only', and coverage-audit.md states 'contains only Markdown files', although icon.jpeg, 01-use-cases.jpg, and 02-workflow-orchestration-schema.svg are in the repository — this does not affect instructive content, but the statement is literally false.
- What is the Agents Best Practices SkillProof Score?
- 9.2/10 — installs cleanly 5/5, triggers reliably 5/5, output vs. baseline 9/10, docs & honesty 4/5.
- How do I install Agents Best Practices?
- Copy the install command from this page, run it in your terminal, and restart Claude Code. Skills live in ~/.claude/skills/ (global) or .claude/skills/ inside a project.
- Can I use Agents Best Practices with Cursor, Copilot, Gemini CLI, Codex or other AI tools?
- The SKILL.md format is native to Claude (Claude Code, Desktop, claude.ai). The instructions inside adapt to other assistants: Cursor rules, GitHub Copilot instructions, Windsurf rules, Custom GPTs, AGENTS.md for OpenAI Codex, and GEMINI.md for Google Gemini CLI — our conversion guides cover each, and the free converter on the tools page does the wrapping for you.