12

Many hands

Module 4 · Operations. After this lesson you can: run multiple AI workers in parallel without collisions, and close a work session cleanly with a handover that keeps the next start fresh.

Running two or more AI workers at the same time is genuinely useful. Two pieces of work advancing simultaneously, without waiting on each other, can produce real time savings. The point where this goes wrong is predictable: two workers in the same workspace, on two unrelated jobs. One finishes first and, being helpful, gathers everything in the shared workspace, including the other worker's half-finished output, and folds all of it into its own result. Two clean work streams go in. One tangled knot comes out. Undoing the tangle takes longer than the parallelization saved.

Neither worker malfunctioned. The collision was built into the setup the moment two workers shared one workspace. Apply the root-cause question from Lesson 6: shared space plus simultaneous hands equals collision, eventually, always. The same dynamic occurs in any physical workspace. Two cooks, one cutting board, one knife: the problem is not the cooks. The structure is the problem.

The fix is structural, not behavioral. Each job run in parallel gets its own lane: its own copy of whatever it is working on, its own space, untouched by anyone else until it is finished. Finished work flows to one place, the operator, who merges results one at a time, checking each against its definition of done from Lesson 9. One merge point, under the operator's control. When one job needs information from another, the answer is: it waits. The first piece lands, gets checked, becomes part of the settled record, and the second job starts fresh from there. Work flows to the center, never sideways. Sideways is where collisions come from.

Which tasks deserve parallel lanes at all? Independent ones. Two tasks are independent when neither needs the other's output and they touch different things. Researching venues while drafting a separate communication: independent. Drafting a message while another worker rewrites the same message's tone: that is one job wearing two names, and running it in parallel produces a race, not speed. The test is simple: ask what is actually saved. Parallel lanes only pay when the jobs genuinely cannot collide.

One further habit: when a phase of work ends, end the session with it. A long conversation carries its whole messy history forward, including everything already settled or superseded, and Lesson 0 covered what happens when the context window gets crowded. A fresh start with a clear handover, what is decided and what comes next, beats a long dragging history every time. Separate stations, one counter where results come together, and a clean reset between phases. That is the structure of a workflow that holds.

Stop here. Think of two tasks you would love to run at the same time.

Can they collide? Touch the same thing, or need each other's results?

Try this nowUnder 30 minutes
  1. Pick two genuinely independent tasks from your week. Different topics, different materials, neither needs the other. Write a done sentence for each.
  2. Run them at the same time, in separate lanes. Two separate conversations or sessions, each with only its own job. Do not mention one job inside the other.
  3. Merge at your counter. When both finish, check each against its own done sentence, one at a time. You are the only place the two results ever meet.
  4. Notice the pull to cross-contaminate. Somewhere in the run you will want one helper to know what the other did. Write that urge down instead of acting on it. That note is your next sequential task, after both lanes land.

Two questions before you go

Answer, then say whether you were sure or guessing. Being honest about which is the skill being trained.

Where you are now

Module 4 covered the operator's toolkit: define done before every handover, keep the rehearsal and re-run it whenever anything changes, size the tool and the attention to the cost of being wrong, and run parallel lanes without letting work cross sideways.

That toolkit only becomes yours through use. Run something real with it this week. Then go to the Shelf, take the templates, and apply the rule from Lesson 8: teach someone. Explaining what you know to another person is the most reliable test of whether you actually own it.