Skip to main content
This walkthrough builds one agent end to end: a contract reviewer that extracts governing law, indemnification and termination clauses from vendor contracts. Everything shown is from a real session.
cd into your project and run helix — you are at the prompt. Not installed yet? See Install Helix.
Driving Helix from a script or a coding agent? Send each command with helix -p "<command>" and continue the same session with helix -c -p "<next command>". Without the interactive screen Helix cannot ask its interview questions, so give the answers in the /spec prompt: the output format, what to do when a clause is missing, and where the agent runs.

​
Spec the contract reviewer

One sentence starts the interview:
Helix reads your project for context (a sample contract in contracts/ gets noticed and used), then interviews you — structured questions with written-out options, not a blank form:
An animated capture of the spec interview: the dashboard, the /spec command being typed, the agent reasoning about the contract-review domain, and the structured question dialog with tabs for The pain, Who uses it, Output shape and Missing clauses.

A real /spec run, trimmed and sped up — about 8 seconds of a 3-minute session.

For this agent it asked, among others: The interview ends with a validated agentspec.yaml — the portable definition of the agent. An excerpt of what this run produced:
.mutagent/specs/contract-reviewer/agentspec.yaml
The spec also carries the agent’s system prompt, its standard operating procedure (ingest → locate clauses → structure the output), and — this matters for the rest of the loop — its evaluation contract: verbatim-extraction-accuracy, missing-clause-detection, structured-json-validity. The session closes by naming the next step:

​
Build it

Build implements the validated spec into the target the interview chose — here, a markdown agent at .claude/agents/contract-reviewer.md — through an implement-and-verify loop, and stops at a working agent. Build has the mechanism.

​
Evaluate it

Evaluate scores the agent against the criteria the spec declared: did every citation match verbatim contract text, did silent clauses get flagged rather than invented, did the JSON validate. Each criterion returns pass/fail with a confidence, rolled into one verdict. Evaluate has the full pipeline.

​
Improve one you already have

You don’t have to start at Spec. Point Helix at an agent that already runs and has traces:
Where it flags failures, ask why:
Diagnose returns ranked remedies. Nothing is applied until you approve. Approve one in Diagnose to apply it once, or hand it to Optimize, which applies it and re-evaluates until your criteria pass.

​
You’re done when

  • The Helix dashboard renders (/help repaints it any time), and
  • /spec (or just asking for a spec) starts the interview, and
  • after the spec and build, .mutagent/specs/contract-reviewer/agentspec.yaml and the agent file the interview chose (here .claude/agents/contract-reviewer.md) exist in your project.
A full /evaluate needs an agent and its traces. A new agent has no traces yet: run it on a few sample contracts first (for example, ask Claude Code to use the contract-reviewer agent on them), then /evaluate contract-reviewer scores those runs.
Prefer prose to commands? Every command here has a natural-language equivalent — “I want a contract-review agent…” routes to /spec the same way. See Commands.