Agent Loop

What It Means For End Users

YeePilot is not limited to one prompt and one command. For larger tasks it can work through a sequence:

1. understand the goal
2. gather context
3. propose a plan
4. execute approved actions
5. verify the result
6. recover from failures when safe to do so
7. finish with a summary

This staged behavior is what the docs call the agent loop.

When You Notice It

You are most likely to see the agent loop when you:

• use /plan <task>
• ask YeePilot to work through several files
• ask for troubleshooting, repair, or migration help
• run a task that needs verification after changes

The interactive client already uses the modern agentic runtime by default. You do not need to turn on a separate "agent mode" for normal use.

The Most Important User Control: /plan

Use plan mode when you want YeePilot to stay explicit and structured:

/plan on
/plan status
/plan Review this repo and propose the safest way to add health checks

/plan on makes plan-first behavior persistent for the session. /plan <task> is the best choice when you only want one task handled this way.

Autonomy Profiles

Autonomy controls how aggressively YeePilot proceeds after planning:

/autonomy status
/autonomy guarded
/autonomy strict_review

Available profiles:

• high
• guarded
• strict_review
• ultra

Recommended for most end users:

• guarded for normal work
• strict_review for production-sensitive tasks

Verification And Recovery

When YeePilot changes files or runs a mutating workflow, it tries to end with evidence that the task actually worked. That evidence can include:

• a clean command result
• a successful build
• a successful test run
• a service or health-check result

If verification fails, YeePilot can run bounded recovery attempts instead of looping forever. Recovery is intentionally limited by budgets in your configuration.

Vibe Coding And Local Mutation Access

Some plans or subscriptions may allow local coding and repository mutation tools, and some may not. YeePilot exposes that as vibe_coding.

For end users, the important rule is simple:

• when vibe_coding is allowed and enabled locally, YeePilot can use local coding and file-mutation tools
• when it is off locally or blocked by the active plan, local coding-agent tools stay disabled

You can see or change the local toggle in the HUD Settings tab when your current plan allows it.

Best Way To Use The Agent Loop

For important work:

1. start with /plan <task>
2. keep autonomy at guarded or strict_review
3. attach the relevant files first
4. watch verification results before you accept the task as done
5. use /export when you want a shareable record

This gives you the benefits of multi-step assistance without giving up control.