agentsmith

AgentSmith — the open-source harness for coding agents

Proof before done.

Move from vibe coding to agentic engineering. Give Claude Code or Codex your project rules, a definition of done, and checks that produce evidence. Understand the task, make a bounded change, verify the result, and carry decisions into the next session.

Project rules and configured guardrails help address scope drift, repeated repair loops, and “done” claims without evidence.

• Start with the right problem. Read existing code and define completion checks before editing.
• Keep changes reviewable. Make small changes, choose native permissions, and enable optional secret-scanning hooks.
• Prove the result. Run all configured checks, exercise the actual feature, and retain the evidence.
• Carry progress forward. Save decisions and a clear handoff in your project.
• Improve from experience. Capture a failure, propose a rule, hook, or test, and validate the change.

Try the difference on one task. Run the disposable demo: reproduce a failure, make a bounded fix, and inspect the proof. Or inspect the existing evidence bundle, including the failing test, verification, real-path output, receipt, and handoff/resume.

Once your checks are trustworthy, explore finite maker/checker runs with approved scope and explicit limits.

Set up AgentSmith

Choose one route. Both use the same guided setup and produce the same project files.

1. Manual setup

1. Download the standalone build for macOS, Windows, or Linux from the latest release. The macOS and Windows builds are currently unsigned, so your operating system may ask you to confirm before running them.
2. Open a terminal and run agentsmith.
3. Answer the guided questions. AgentSmith shows the exact plan before it writes anything.

The wizard can configure an existing project or create an empty folder and initialize Git for a new one. The standalone app includes its runtime; Python is not required.

2. Ask your agent

Paste this into Claude Code, Codex, or another coding agent:

Install AgentSmith for this project. Follow the official agent guide at
https://github.com/PromptPartner/agentsmith/blob/master/AGENT-INSTALL.md
Inspect the project first, explain your profile recommendation in plain language,
show me the exact installation plan, and ask once before applying it.

The agent inspects the repository without running project code, recommends a profile from evidence, previews the managed files, installs AgentSmith, and runs the doctor check.

Building AgentSmith from source is for contributors and advanced automation. See the installation reference.

New project or existing project?

Project setup is the default because the rules travel with the repository and can be reviewed by collaborators. A user-wide core for every project is available under Advanced options. A layered setup combines that user-wide core with a small project-specific profile. See project, user-wide, and layered setups.

Work profiles

A profile tells the agent what “done” means for the work at hand. The wizard recommends one profile and shows the file evidence behind its choice. You can accept it or view the full list.

autonomous-loops is an advanced modifier for scheduled or unattended work. Combine it with the main work profile; do not use it by itself. The profile guide explains close calls, switching, and stacking.

What gets installed

• AGENTS.md is the canonical project agreement. Claude Code also receives a generated CLAUDE.md.
• .harness/verify.conf defines the real checks for this project.
• .agentsmith/state.json records only AgentSmith-owned settings so updates and removal preserve foreign content.
• Optional skills, MCP servers, and hooks are available in the collapsed advanced step.

The cautious permission mode is the default. AgentSmith never treats a connected external service as permission to write to it.

See the proof

The First Verified Loop records a failing test before the fix, passing verification afterward, a real command invocation, a receipt, and a resumable handoff. The support registry separates instruction support from tested native behavior.

Documentation and community

• Installation reference
• Documentation map
• How verification becomes evidence
• Agent compatibility
• Contributing and support
• Security policy and code of conduct

MIT licensed. Built by PromptPartner.

Optional integrations disclose their effects before installation. See capability selection.

Recall local notes with agentsmith memory search "topic" or agentsmith memory read docs/research/note.md. Optional --with-memory-startup adds reference-only startup suggestions; see project memory. Orca delivery requires host observation.

agentsmith audit-config provides a read-only advisory report for selected project configuration. --include-user adds native user settings; --fail-on high adds a CI gate. See configuration audit for scope and limits.