From agent-harness
Building or reviewing agent-harness code — the agentic loop, tool design, context and memory management, system prompts, subagents and multi-agent orchestration, permissions/sandboxing and safety, lifecycle hooks, error handling and resilience, streaming, observability, or evaluation — and decisions like workflow-vs-agent and single-vs-multi-agent. Use when writing, debugging, or reviewing the runtime loop around an LLM (especially with the Claude Agent SDK).
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-harness:agent-harnessThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
A fact-checked reference on building agent **harnesses** (the runtime/agentic loop that
A fact-checked reference on building agent harnesses (the runtime/agentic loop that wraps an LLM), bundled with this plugin. It is built from 25 primary sources and 150 adversarially-verified claims. Don't dump the whole guide into context — route to the one relevant section and read only that.
Route first. Read the navigation manifest — it pre-joins each guide section
(stable ids #s0–#s12) to its code samples and primary sources, so you can find
what backs a topic in one lookup:
${CLAUDE_PLUGIN_ROOT}/index.json
Pick the section whose title matches the task; note its lines range and samples.
Read just that section (use the lines range — do not load the whole file):
${CLAUDE_PLUGIN_ROOT}/docs/agent-harness-best-practices.md
Each section is self-contained: principle → concrete practices → inline citations →
a → Code: pointer to its runnable sample.
Open the sample the section points to, under ${CLAUDE_PLUGIN_ROOT}/code_samples/
(TypeScript is primary; Python parallels exist for the first three themes).
For provenance, resolve a citation key (e.g. [be], [wt], [ce]) to its source
via ${CLAUDE_PLUGIN_ROOT}/raw_outputs/sources/INDEX.md. Per-claim evidence (the exact
supporting quote and verification votes) is in
${CLAUDE_PLUGIN_ROOT}/raw_outputs/claims/verified-claims.json — a heavy file, so load
it only when you specifically need claim-level provenance.
For fast "what should I do here?" decisions, jump straight to the Quick decision cheat-sheet and Top anti-patterns near the end of the guide.
Fetch the same files from the public repo with WebFetch instead:
https://raw.githubusercontent.com/blixxurd/context-agent-harnesses/main/index.jsonhttps://raw.githubusercontent.com/blixxurd/context-agent-harnesses/main/docs/agent-harness-best-practices.mdhttps://raw.githubusercontent.com/blixxurd/context-agent-harnesses/main/raw_outputs/sources/INDEX.mdGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub blixxurd/fidget-marketplace --plugin agent-harness