Post

Here's the latest approach I've been taking to autonomous work that gets authorized and queued up for my agents to work on.

1. Every significant piece of work starts with a PLAN file that captures the big picture -- what's different after it's done, why would I stop doing that work, etc...

The goal is to get broad alignment on what we’re aiming for. And, if the plan isn't marked as aligned, the next step can't run and we can’t move forward.

2. Next, a skill called "decompose" breaks the plan down into appropriate steps, each with its own guardrails and parameters like what tools it can use, folders it can work in, etc.

3. Then, an independent reviewer agent checks the plan and decomposition without access to the model's reasoning (so it isn't just agreeing with what it previous decided!).

4. A skill called simplify then runs with one goal: make the plan smaller while still meeting the goal. (The previous step often introduces more steps in an effort to “fix” what it found) This skill can remove steps, merge or simplify, but it CANNOT add any steps.

This has been running for 7 days (yes, I always seem to make big changes to my systems on Friday night). Some stats... in 19 plans over this past week: - 100 of 277 (36%) steps were marked to run on their own

- one plan went from 7 steps to 26 (from the adversarial reviews and fixes) and then the simplify skill brought it back down to 17 - 54 of the 118 steps that finished this week ran completely on their own, end to end.

Longest streak of autonomous tasks done in a row: 12 steps... about 82 minutes work, completed in 26 minutes on the clock because they were able to run in parallel.