Documentation / The agent

Missions and phases

Every large job is a mission with an 8-phase lifecycle, an execution plan, and a goal that never gets lost.

Every large job is a mission. It's saved with the project and keeps the original request, the requirements, the acceptance criteria, the decisions, the assumptions, the plan, the files touched, and where the work stands.

That way the goal and the plan are never lost, no matter how long the conversation grows.

The 8 phases

# Phase What it answers
1 Analysis What exactly is being asked? What is still unknown?
2 Design How will it be solved? Which decisions are made?
3 Planning Which verifiable steps does it break down into?
4 Implementation Write the code, touching only what's necessary.
5 Measurement Is the touched code better than the previous version?
6 Testing Build, test, and start the application.
7 Review Review its own diff like a strict reviewer.
8 Delivery What was done, what was verified, the decisions, and what's pending.

Measurement compares the touched code with the previous version (N−1) and sends the work back to Implementation until it's better. It can't be skipped. See Measurement against N−1.

Phases follow the work

  • Every new request starts in Analysis.
  • The phase advances based on what the agent does, not just on what it declares: reading and searching is Analysis, editing is Implementation, running tests is Testing.
  • It only moves forward. Reading a file while you implement is still implementing.
  • It works the same with Yuli's agent, Claude Code, Kiro, Codex, and Antigravity.

The mission bar

It appears above the chat, below the conversation tabs.

  • The eight stops, with completed ones checked and the current one highlighted.
  • Phase n of 8 · Name, since when, and what it's doing.
  • The Mission trail, with the time spent in each phase.
  • The current measurement, with the See in Quality button.

When you expand it, you'll see:

Section Contents
Execution plan The steps, Step n of N, the estimated time, and who runs each step.
Requirements What must be met.
Acceptance criteria How you know it's done.
Recent decisions What was decided and why.
Unattended The unattended mode switch.

Actions: View plan, Close mission, and Reopen mission.

Remove a plan

The ✕ button asks “What should I do with this plan?”:

  • Cancel plan: stops it; the history (PLAN.md) is kept.
  • Delete: deletes the plan, the log, and the mission from this conversation.
  • Keep: leaves it as it is.

The workboard

The workboard shows live what the agent and its team are doing during a turn.

  • States: Analyzing, Investigating, Editing, Verifying, Running plan, Waiting.
  • Agents: the Lead agent and each Subagent.
  • View full work opens “Turn work” with the Timeline, Agents, Plan, and Team tabs.

In Subscription mode, your agent's plan (for example, Claude Code's task list) is mirrored on Yuli's workboard.

No cutoffs from round limits

If a turn runs out of rounds, the agent leaves a handoff point and the next turn continues from there. In PLAN.md it appears as “Where the work stands”. Context stays bounded through incremental summaries.

Where it's stored

.ai/missions/<conversación>/
├── mission.json   the complete mission
└── PLAN.md        the readable plan

There is one mission per conversation. See The .ai folder.

SDLC documentation

Tools → Lifecycle documentation (SDLC)… generates the project's development lifecycle documentation.