Active Loop

An iterative improvement protocol for agents. The agent declares a measurable optimization target, runs repeated cycles via /loop, experiments with different approaches, measures outcomes, and accumulates all work in a DRAFT PR with a learning log.

When to Use

When a task requires repeated, incremental improvement over multiple cycles rather than a single execution pass. Examples:

• Task graph densification and maintenance
• Code quality improvement across a large codebase
• Data cleanup or migration in batches
• Any work where "try, measure, learn, repeat" beats "plan everything upfront"

Protocol

Step 0: Declare Optimization Target (MANDATORY)

Before any work begins, the agent MUST:

1. State the target in measurable terms (e.g., "reduce flat tasks from 226 to <100")
2. Define the measurement — which tool/command produces the metric
3. Confirm with the user — do not proceed until the target is acknowledged

If the target is unclear, ask the user. Do not guess.

Optimization target: [what we're improving]
Metric: [how we measure it]
Baseline: [current value]
Tool: [command/tool that produces the metric]

Step 1: Initialize (First Cycle Only)

1. Take baseline measurement using the declared tool
2. Create a working branch: active-loop/<target-slug>
3. Open a DRAFT PR with the optimization target and baseline in the body
4. Record state in the PR body (see PR Body Format below)

Step 2: Each Cycle

Every cycle follows this sequence:

MEASURE → DECIDE → DO → MEASURE → LOG → PUSH

1. Measure current state — run the metric tool, compare to last cycle
2. Decide approach — based on what we learned last cycle, pick what to try this time. Rotate strategies. Don't repeat what didn't work.
3. Do the work — bounded effort per cycle (e.g., 10 tasks, 20 minutes, one batch)
4. Measure again — same tool, capture delta
5. Log learnings — what we tried, what changed, what surprised us, what to try next
6. Commit and push — descriptive commit message referencing the cycle number
7. Update DRAFT PR body — append cycle log entry

Step 3: Re-confirm Target

After each cycle, check:

• Is the target still the right target?
• Are we making progress?
• Should we adjust the approach or the target itself?

If running via /loop, output a brief status line so the user can see progress without interrupting.

Step 4: Completion

When the target is met (or the user decides to stop):

1. Final measurement
2. Summary of all cycles and total improvement
3. Mark PR as ready for review (remove DRAFT)
4. Complete the tracking task

PR Body Format

## Optimization Target

**Target**: [what]
**Metric**: [how measured]
**Baseline**: [starting value]
**Goal**: [target value or "continuous improvement"]

## Cycle Log

### Cycle 1 — [date]

- **Approach**: [what was tried]
- **Before**: [metric value]
- **After**: [metric value]
- **Delta**: [change]
- **Learnings**: [what we discovered]
- **Next**: [what to try next cycle]

### Cycle 2 — [date]

...

Integration with /loop

To run an active loop on a schedule:

/loop 30m /active-loop <task-id>

The agent reads the PR body to recover state, runs one cycle, updates the PR, and exits. The /loop command handles re-invocation.

When resuming from a previous cycle:

1. Find the DRAFT PR for the branch active-loop/<slug>
2. Read the cycle log from the PR body
3. Parse the last cycle's "Next" field to inform this cycle's approach
4. Continue from Step 2

Anti-patterns

• Starting work without declaring the optimization target
• Running cycles without measuring before AND after
• Repeating the same approach when it didn't produce results
• Making the PR body so long it becomes unreadable (summarize older cycles)
• Skipping the "learnings" field — this is the whole point
• Changing the optimization target without user confirmation

Composability

Active-loop is a protocol, not a domain skill. It composes with domain skills:

Domain	Compose with
Task graph health	[[planner]] (maintain mode), graph_stats
Code quality	linters, test coverage tools
Knowledge base	[[planner]] (maintain mode), PKB search quality metrics
Framework governance	[[audit]], compliance metrics