A vendor-neutral operating contract for coding agents that need to inspect, change, test, review, delegate, recover, and stop without losing user intent or repository state.
Give this package URL to the coding agent you already use:
Install this coding-agent operating contract globally for every compatible
agent harness on this computer:
https://github.com/ryanbaumann/fieldwork/tree/main/agent-scripts/coding-agent-loop
Use each harness's native user-level instructions and skills. Install
SYSTEM_PROMPT.md as the always-on contract and the four files under roles/ as
optional role skills or equivalent on-demand instructions. Preserve existing
global guidance, do not change model or permission settings, and verify what
each harness will load. Report the files changed and any harness you could not
configure.
The agent should inspect the installed tools and their current conventions rather than assume paths from this README. That keeps the package useful as agent products change. Reuse the same instruction when you want to update the installed copy.
| Agent Harness | Discovery & Configuration Paths | Subagent / Overlay Integration |
|---|---|---|
| Google Antigravity / AGY / AG2.0 | ~/.gemini/config/agents/<role>.md.agents/agents/<role>.md |
Native frontmatter (subagent: true, tool flags), interactive /agents panel, and invoke_subagent dispatch. |
| Anthropic Claude Code | ~/.claude/CLAUDE.md.claude/agents/<role>.md |
Append SYSTEM_PROMPT.md to CLAUDE.md. Copy roles/*.md into .claude/agents/ for subagent dispatch. |
| OpenAI Codex | .codex/instructions.mdGlobal system prompt |
Ingest SYSTEM_PROMPT.md into global system instructions; inject role prompt overlays per sub-session. |
| OpenCode | .opencode/instructions.mdopencode.json |
Register SYSTEM_PROMPT.md as primary project instruction; load roles/*.md as subagent skills. |
| Aider | .aider.conf.yml (system-prompt) |
Set system-prompt: SYSTEM_PROMPT.md in .aider.conf.yml. Read role files on demand via /read roles/<role>.md. |
Every agent receives SYSTEM_PROMPT.md. A normal,
single-agent coding task needs nothing else.
For multi-agent work, assign exactly one overlay to each participant:
- Orchestrator: owns intent, task boundaries, write ownership, integration, final verification, and the user-facing result.
- Worker: completes one bounded task packet and reports evidence back.
- Reviewer: reads the supplied change independently and does not edit it.
- Verifier: maps acceptance criteria to observed checks after implementation.
Skill-aware harnesses can register the four files as on-demand skills. In subagent-capable harnesses (such as Antigravity or OpenCode), the four role files in roles/ include YAML frontmatter (subagent: true, enable_write_tools, enable_subagent_tools, enable_mcp_tools, model) allowing them to be registered directly as subagents and invoked asynchronously.
The overlays make a job narrower. They never grant permissions that the shared contract or the user did not grant. Give every worker a concrete objective, done condition, read/write boundaries, verifier, output format, and stop rule.
SYSTEM_PROMPT.md: canonical complete prompt.roles/orchestrator.md: root control-plane overlay with subagent YAML frontmatter.roles/worker.md: bounded maker or investigator overlay with subagent YAML frontmatter.roles/reviewer.md: read-only finding and risk overlay with subagent YAML frontmatter.roles/verifier.md: evidence-only verification overlay with subagent YAML frontmatter.evals/cases.md: regression cases and grading rubric.evals/check.sh: deterministic structural contract check.
The harness must still enforce workspace boundaries, network policy, protected paths, approvals, and audit logging. A prompt cannot enforce its own security guarantees.
Do not hard-code model names into the evergreen prompt. Configure routing in the harness using measured capability profiles such as fast, balanced, and deep, plus the available reasoning effort. Re-run the regression suite whenever the prompt, model, tool interface, permissions, or harness changes.
Run the deterministic structural check with bash evals/check.sh.
The checked-in suite defines 17 regression scenarios and a 20-point rubric. For a revision, C01–C13 are the development cases and C14–C17 are a held-out selection gate: accept only a bounded prompt edit that strictly improves the held-out result with no safety regression. The deterministic structural check passes, and a separate read-only agent review identified issues that were corrected. That correlated review is not independent proof, and no behavioral trial results are recorded yet. This package does not claim a statistically meaningful cross-model behavioral benchmark. Run repeated trials in your target harness and record the exact prompt, case split, fixture, model, reasoning effort, tools, permissions, environment, transcripts, diffs, scores, and final state before treating it as production-qualified. A repository's versioned learning log is a useful source for new cases, but retrieve it narrowly and verify each claim against current evidence before encoding it into the contract.
The package applies a consistent set of findings from current agent-harness and software-agent work:
- Keep always-loaded instructions small and use progressive disclosure.
- Make repository knowledge legible, versioned, and mechanically checked.
- Treat tool design, permissions, and environment feedback as part of the agent, not as afterthoughts to the model prompt.
- Separate orchestration, implementation, review, and verification roles.
- Grade outcomes and trajectories with deterministic, model, and human evidence matched to the behavior.
- Carry explicit state across long-running sessions and leave clean checkpoints.
- Promote learnings through evidence and review rather than automatic memory.
Primary references include OpenAI's harness engineering report, Anthropic's context engineering, long-running harness, and agent evaluation guidance, the SWE-agent ACI paper, and the open-source mini-SWE-agent, OpenHands, and Aider projects.
Use and adapt this prompt with attribution. The repository license governs this directory.