ReachmadeLab
All productsContact
Product 05 / 07MULTI-AGENT SYSTEMSUnder evaluation · MIT · repository: Multibot

Agent Team

Ask once. Keep only the version that was checked.

AI teammates share one request. A teammate other than the maker checks each revision, and you keep which version passed what — plus whatever was left unfinished.

Watch the filmDiscuss this technology
A frame from the design film (recreated screens, partly not implemented)
FIG. 05A frame from the design film (recreated screens, partly not implemented)

What it does

  1. Request
  2. draft
  3. check
  4. checked version
(03) Agent Team — FilmDesign film (30 s · silent · recreated screens, partly not implemented) Rendered from design files; not a recording of the real app. The request, conversation and checks are fictional examples. Watch a recording of a real run ↗

(+) In detail

05 / MULTI-AGENT SYSTEMS / Under evaluation

AI agents that draft, review and revise

Give that draft a second pair of eyes.

One AI agent writes. Another checks the draft against your source. Findings go back for revision, and the code or document keeps a record of what was checked in each version.

Read a draft-to-revision example Inspect the roles and code

The public page is a record of a real-model run. Clicking it does not start a new model run in the browser.

Try the guided sample first Sample only. No AI run or external action. Editing needs JavaScript.
Agent TeamREAL PRODUCT / ~13 SEC / 1×
Agent work trail
Real-model work records and evaluation results including failures. The featured research replay ended partial. This site edit is about 13 seconds; playback speed is unchanged.Source recording

FIRST THING TO TRYRead a draft → review → revision trail, including what remained unfinished.

Setup, scope & license

MIT · repository: Multibot · A quality advantage over a single agent has not been established. Published comparisons include failures and provider constraints. Production acceptance is separate.

  • Not the makerThe Reviewer is separate from the maker; findings are fixed in the next revision
  • Per-revision checksChecks bind to one revision’s SHA-256; a new revision starts unchecked
  • Saved with its recordSave chosen revisions as a ZIP with a SHA-256 manifest, report and events

WHAT THE FILM SHOWS

See what changed after the review.

  1. Ask

    Describe the goal and the artifact, and attach your source. The coordinator sets the plan and the finish conditions.

  2. Made, then checked by someone else

    The maker publishes a revision and the reviewer checks that revision condition by condition. Findings are fixed in the next revision.

  3. Keep the checked version

    Save the chosen revision with its record, including which checks it passed and what remained unfinished.

TRY IT

Read the before. Then the after.

Agent Team separates drafting, review and revision while keeping artifacts connected to the work trail. Public examples retain failures and partial completion; no quality advantage over a single agent is claimed.

Read a draft-to-revision example Inspect on GitHub

CURRENT BOUNDARIES

Availability and development.

Current scope

A quality advantage over a single agent has not been established. Published comparisons include failures and provider constraints. Production acceptance is separate.

MIT · repository: Multibot

Read the evidence

Proposed interface design

A redesign with a sidebar, a workroom that keeps conversation and result side by side, and marked findings and fixes. The frames are renders of a design file, not recordings of the app. The request, messages and checks are fictional, and the stills show the Japanese interface. These design renders do not establish the current implementation status of each feature.

  1. Design render: a sidebar layout with the request, the attached product.md and the “Ask the team” button.
    Design preview · fictional example01 AskOne sentence and a source file are enough.
  2. Design render: the Pon, Mame and Mugi characters hand work over in chat bubbles, with guide.md version 1 open on the right.
    Design preview · fictional example02 Share the workWho handed what to whom, in the conversation.
  3. Design render: one sentence of version 1 is marked in amber with “Lulu’s finding: goes beyond the source”.
    Design preview · fictional example03 FindingThe reviewer’s finding sits on the sentence it is about.
  4. Design render: the fixed sentence is marked in green and four check items show a pass mark.
    Design preview · fictional example04 CheckedFixed in version 2; all four conditions pass.
  5. Design render: a ZIP list with guide.md version 2, manifest.json, report.md and events.jsonl.
    Design preview · fictional example05 SaveThe chosen version, zipped with its check record.
The separate design film (Launchloom)

(05) What it can do

The maker and the checker are different.

Source: Check the README · checked against the repository on 2026-09-27

F-01 / ROLES

The maker is not the checker

A Master sets the deliverables and finish conditions while a Researcher, a Builder and a Reviewer do the work. In My team you can pick names and voices from the Mame, Pon, Mugi and Lulu characters. No quality advantage over a single agent is claimed, and published comparisons include failures.

  • Master / Researcher / Builder / Reviewer
  • Characters are independent of roles
  • Failures published
F-02 / REVIEW

Catch what disagrees with the source

The Reviewer compares the actual revision with cited sources and attachments; anything it cannot fetch stays unverified. Findings go back to the maker and are fixed in the next revision.

  • Source check
  • Unverified stays visible
  • Finding → next revision
F-03 / TRAIL

Nothing is overwritten

Everything goes to an append-only event store, with per-revision artifacts and checks side by side. Chosen revisions can be saved as a ZIP with a SHA-256 manifest, report and events. Partial completion is reported as such.

  • Append-only
  • Per-revision checks
  • ZIP with manifest

(06) Scope and evidence

What it cannot do,
stated up front.

Public sources checked 2026-09-15. Check each repository and product page for current terms.

SCOPE

A quality advantage over a single agent has not been established. Published comparisons include failures and provider constraints. Production acceptance is separate.

PROOF / PUBLISHED EVIDENCE

Real-model work records and evaluation results including failures. The featured research replay ended partial.

(07) Build with it

Prototype agent workflows with explicit roles, artifact checks, budgets, permissions, and human approval.

Discuss a project