agent-harness
Design or review an AI agent harness: context assembly and compression, perception tools, action tools, constraints, verification, correction, and evals. Use for agent runtimes, tool loops, coding agents, RAG or document agents, long-running workflows, and context or tool failures; not ordinary one-shot prompting.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 5
- Forks
- 3
01 前言 Front Matter
- name
- agent-harness
- description
- Design or review an AI agent harness: context assembly and compression, perception tools, action tools, constraints, verification, correction, and evals. Use for agent runtimes, tool loops, coding agents, RAG or document agents, long-running workflows, and context or tool failures; not ordinary one-shot prompting.
02 资源文件 3
- 参考文档agents/openai.yaml190 B
- 参考文档references/context-and-perception.md2.6 KB
- 参考文档references/evaluation-and-correction.md1.8 KB
03 SKILL.md 正文
Agent Harness
Treat the harness as the runtime system around the model. Its job is to control what the model can perceive, which actions it can take, and how results are checked and corrected.
Workflow
- Define the task boundary, observable success, stop conditions, and actions requiring human authority. Finish when success and failure are measurable.
- Map one decision loop: assemble context, decide, act, observe, verify, recover. Name the owner and output contract for every stage.
- Inventory the agent's eyes: instructions, state, history, memory, retrieved evidence, tool schemas and output, errors, and environment feedback. Read context and perception when designing or debugging these inputs.
- Design narrow tools around decisions. Prefer candidate lists followed by bounded detail reads, stable identifiers, pagination, explicit truncation, and structured errors. Separate read-only perception from mutation.
- Add constraints before actions, verification after actions, and a bounded correction path after failures. Make irreversible actions explicit.
- Build the evaluation before broadening capability. Read evaluation and correction for the boundary, retention, trajectory, and anti-gaming checks.
- Implement the smallest complete loop, record its inputs and outputs, run the focused evals, and inspect failures. Expand only when the trace shows a missing capability rather than a prompting preference.
Non-negotiables
- Keep stable instructions and tool definitions in a reusable prefix. Append dynamic task evidence and state in deterministic order.
- Preserve decisions, constraints, failures, evidence identifiers, and current state during compression. Never hide truncation.
- Prefer text and structure for semantics; add images or screenshots when layout, geometry, or rendering is part of the decision.
- Keep the evaluator independent from the producer when practical.
- Retain unsuccessful trajectories. They explain whether the defect was in perception, reasoning, action, verification, or recovery.
04 同仓库更多 Skills
potetomode
Load pstack's poteto-mode workflow for concise responses, deliberate delegation, simple code, and verified work. Use when the user requests potetomode or /potetomode.
bro
Restate the last assistant message in plain, concise language using pstack's bro skill. Use when the user requests bro or /bro.
wayfinder
Plan a huge chunk of work (more than one agent session can hold) as a shared map of decision tickets on your issue tracker, and resolve them one at a time until the way to the destination is clear.
using-exe-dev
Operate exe.dev virtual machines. Use when the task mentions exe.dev, exe VMs, *.exe.xyz hosts, or exe.dev infrastructure.
unlazy
Prove substantial work complete with file-backed gates and runnable checks. Use for exhaustive completion, a Depth Tree, or a completion ledger.
sync-skills
Sync vendored third-party skills against their upstream repositories, re-apply local patches, and flag new upstream skills. Invoke explicitly with /sync-skills.