Misfit Circuit

The Apartment · private · runs locally

One request. A team of AI agents from rival companies.

The Apartment is my local agent setup, where the agents do their work. Composer is the app that runs a team of them on one project. You start from a workflow you can read, then look at or change the steps, models, roles, order and stopping points. Each step passes its evidence to the next, and the builder never grades its own work.

A run can build and review without installing anything. Installing it or changing an outside service is a separate, explicitly authorized step.

See the setup steps Follow a recorded run
Illustration: the four crew robots and Steve, seen from behind, around a round planning table covered in maps, models, and pinned notes; through the window, the lit hilltop town.
Composing a plan together at the round table. Original illustration for the Misfit Circuit project · not a product capture.

What it is

One job, split into rooms.

Composer gives different parts of a job their own agents, so the builder doesn’t grade its own work. The work is split into rooms, each with a clear job.

  • Pick a workflowYou pick how to begin: Investigate, Build or Review. Composer suggests a crew that fits, and you can change any of it.
  • One job per roomBuilders write the code. Reviewers read and report, and change nothing. Each agent works in its own space.
  • Evidence moves forwardLater rooms read only the sealed outputs of earlier ones. When the run ends, Composer writes a record of what happened.

One room investigates, one builds, one tries to break what got built. The main crew sets up the run and keeps an eye on it. The workers inside are brought in for a single task, and the run record notes each worker’s model family.

The Composer

Workflow first. Advanced controls second.

Composer is installed locally, and its interface passed desktop and phone-width browser checks on October 3. The queue driver stays off. The design preview below uses sample data and launches nothing. The step-by-step screens after it are earlier pre-install captures from a scratch server with synthetic data.

Illustration of the Composer’s entrance, drawn for this page from the Composer’s screens · not a product capture

Step 1
How should we begin?
Not available yetCompare approachesIndependent takes before choosing. No workflow runs and compares approaches yet, so this stays switched off.

Investigate suggests a question crew.

The Investigation Room is a preset here, not a separate app. It scales only when the question earns it.

  • Quick · one seat

    One investigatorA bounded review for a narrow question that does not need independent coverage lanes.

  • Layered · 3 waves

    Five seats across three wavesIndependent concerns are sealed before synthesis; the final reviewer records its own coverage before reading prior reports.

Build suggests a builder, then a reviewer.

Every stage and lane can still change; sealing is a later, separate step.

  • Wave 1 · can start first

    BuildBuilder: writes or changes code and files. Definition of done: the files it must produce.

  • Wave 2 · waits for wave 1

    ReviewReviewer: reads and reports, changes nothing. It reads only the sealed outputs of the build.

Review suggests a reviewer who changes nothing.

For evaluating something that already exists.

  • Wave 1

    ReviewReviewer: reads and reports, changes nothing. Add participants for independent takes; findings are merged by evidence, not by vote.

Step 2 · optional Open advanced composition

Everything the preset suggested can be inspected and changed here before sealing.

  • Models and lanesEach stage names its exact lane: provider, model, and account. Readiness shows per lane, and cost stays “unknown” when it is.
  • WavesStages in the same wave may run in parallel. A later wave waits for the one before it, across an explicit handoff.
  • RolesBuilders write or change code and files; reviewers read and report and change nothing. A foreman appears when a stage has two or more participants.
  • OrderingMove stages up or down, or name what each stage waits for. Reviews read only sealed outputs of earlier stages.
  • Stop conditionsEach stage has a definition of done, the files it must produce, and one to five attempts per participant. If a lane fails, the default is to wait for you: nothing switches on its own.
  • Blind seatsA non-review stage can be marked blind: it receives only its brief and the files and notes, not the goal text or earlier outputs.
Then: review and start

Sealing freezes the plan. Start is its own confirmed step and creates one run of the sealed plan, exactly as frozen. Choosing a workflow only opens the form: nothing starts, nothing is sent anywhere.

The current design

Steps run left to right. Add agents from the shelf at the top, then select any step to set its model, how many agents it gets and its task. Underneath, a panel shows what runs first and what comes next.

Composer’s simulated demo: the first step’s two agent cards marked Working, the next step marked Up next, and a Current and Next panel underneath. Sample data, no real agents.
Simulated demo · design preview · sample data · no real agents or provider calls · one provider name hidden
Composer’s design preview: a two-step workflow laid out left to right, a shelf of agents to add at the top, and settings for the selected step on the right. Sample data.
Building a workflow · Design preview · sample data · nothing here launches real agents · two provider names hidden
In light mode · open the light version
Composer’s design preview in light mode: the same two-step workflow and step settings. Sample data.
Light mode · Design preview · sample data · nothing here launches real agents · two provider names hidden
On a phone · open the phone layout
Composer’s design preview on a phone: the same workflow, scrolled sideways, with First and Next underneath. Sample data.
Phone layout · Design preview · sample data · nothing here launches real agents · one provider name hidden

From goal to an explicit Start, in five steps

  1. 1 · Goal

    Say what the work is and what done looks like.

    You begin with the goal and choose how to start: Investigate, Build, or Review. Choosing only opens the form.

    The Composer’s first screen: How should we begin? with workflow cards.

  2. 2 · Inputs

    Name what the crew may read and touch.

    Allowed files and notes are listed up front. A blind seat can be given only its brief and inputs, not the goal text or earlier outputs.

  3. 3 · Compose the team

    Seats, lanes, and waves you can inspect.

    The preset suggests a crew; every stage, lane, role, and ordering can be changed. Reviewers read and change nothing.

    The Composer’s compose-the-team screen: a two-wave plan of Build then Review with participant cards, not running.

  4. 4 · Review and seal

    Freeze the plan before anything runs.

    Review shows what is frozen and who may run each stage. Sealing locks the plan; it still does not start it.

    The Composer’s review screen after sealing, listing what is frozen and who may run each stage.

  5. 5 · Start, explicitly

    A person presses Start. Nothing dispatches on its own.

    Start is its own confirmed step and creates one run of the sealed plan, exactly as frozen. The queue driver stays off, and this page cannot start anything.

Screens are the same earlier captures shown below, from a scratch server with synthetic data; no live-provider run is shown.

The Composer’s own screens, before install

Composer pre-install capture, step 1: How should we begin? Cards for Explore a question, Build or repair, Review or audit, and a dashed, unavailable Compare approaches card.
Step 1 · choose a workflow · pre-install capture · scratch server · synthetic data · full capture
Step 2 · compose the team · open the full-height capture
Composer pre-install capture, step 2: compose the team. Workflow on the left, a two-wave plan of Build then Review in the middle, and participant cards with lanes on the right, all editable and not running.
Step 2 · compose the team · pre-install capture · scratch server · synthetic data · full capture
Review & start after sealing · open the full-height capture
Composer pre-install capture, review and start after sealing: what is frozen, who may run each stage, and Start as its own separate step.
Review & start after sealing · pre-install capture · scratch server · synthetic data · full capture

The mechanism

Five stages, each with a clear stopping point.

  1. Freeze

    One brief, success criteria, allowed inputs, and a defined finish line.

  2. Separate

    Each agent gets one focused role and its own separate workspace.

  3. Merge

    A coordinator deduplicates findings without silently changing their meaning.

  4. Check

    Targeted tests and independent review examine the assembled result.

  5. Stop

    The run stops where it was allowed to stop. Installing and releasing stay separate steps.

A real recorded run

Seven investigators found 18 issues and test gaps.

Four repair agents fixed 16. Two risks were deferred on purpose. The fixed version passed 86 Python checks and 49 browser checks, then stopped before installing.

  1. Investigate

    Seven reviewers worked separately from the same frozen version. Three requested agents were blocked and never launched.

  2. Merge

    The two coordinators merged the separate findings into 18 issues and test gaps.

  3. Repair

    Four repair agents fixed 16 findings. The other two stayed named and deferred.

  4. Review

    A separate reviewer passed it, with limits. The investigation itself didn’t install anything.

What this doesn’t show

What this run does not prove

  • It does not prove parallel workers were faster than one.
  • Total cost and human intervention were not measured.
  • The frozen screenshot predates the repair.
  • Pass-with-limits is not production approval.

Across 193 runs recorded in the Apartment, Steve’s local agent setup, 125 finished, 41 failed and 27 were partial. The failures stay in the record. The agents still make mistakes; the system is built so mistakes get caught and written down, not to prevent them.

It is a scoped record, not an all-time total. No speedup claim. Cost, total human intervention, and causal benefit were not measured.

All five case studies →

How a project moves

Separate rooms, evidence passed forward.

  1. 1Idea
  2. 2Investigation
  3. 3Coding
  4. 4Testing
  5. 5Record

An idea enters the Investigation room. Independent agents look for missing requirements, hidden blockers, and reusable work. A build packet moves to Coding. A different set of agents tests the result and tries to break it. Composer’s board records what happened, so the next project starts from the record instead of from memory.

  • The Apartment is private and runs locally. It isn’t publicly available or hosted.
  • The agents still make mistakes. The system is built so mistakes get caught and recorded; it doesn’t prevent them.