AI LabConductorLaunch console

Multi-agent runs you can actually trust.

A planner decomposes a goal, tool-using workers run it in parallel, and a critic verifies the result before it ships, every step streamed as a debuggable trace.

Runs in your browser. No sign-up, no API key.

What it does

A small orchestra, one goal at a time.

Conductor turns a goal into a performance: a planner writes the parts, workers play them with real tools, and a critic listens before anything ships.

Personnel

Who plays in every run

  • The Plannerwrites the parts

    Reads the goal, splits it into tasks and decides which of them can share a bar.

    One step per run, 190 to 240 tokens

  • The Workersplay with tools

    Research, Extract, Score, Timeline and the rest. Each takes one task and one tool; workers in the same bar play in parallel.

    15 workers across the 8 goals

  • The Criticlistens before it ships

    Checks the result against the goal. When something is missing it sends the step back, and the revision plays in the next bar.

    Sends work back in goals IV and VII

Instruments

The tools the workers call

  • web_searchResearch

    Finds and reads sources on the open web.

    Called in 2 of the 8 goals

  • structured_outputExtract, Tag, Triage

    Normalizes what was found into tables and typed fields.

    Called in 4 of the 8 goals

  • run_code9 workers

    Executes analysis, tests and scans in a sandbox.

    Called in 7 of the 8 goals by Score, Timeline, Analyze, Diagnose, Budget, Cluster, Test, Implement, Scan

  • generateContent, Roles, Implement

    Drafts prose, plans and code.

    Called in 3 of the 8 goals

How it works

Three movements, one verified result.

The score above is the run itself: the same goals, bars and token counts the console plays, timed by the same engine.

  1. Iplan210A

    Plan

    The planner reads the goal and writes the parts: who plays, in what order, and which tasks share a bar. The plan is pencilled in before any worker starts.

  2. IIweb_searchstructured_output

    Perform

    Workers play their bars with real tools: web search, structured extraction, code execution and generation. Workers in the same bar run in parallel and finish together.

  3. IIISENT BACKrev.PASS

    Verify

    The critic listens before anything ships. A gap gets the step sent back, the revision plays in the next bar, and the run ends only when the critic signs.

Tempo moderato
The first bar enters at 0.42 s. Each bar plays 1.00 s and rests 0.38 s, so goal I, in 4 bars, plays in 5.94 s.
Da capo
In goals IV and VII the critic stamps SENT BACK; the worker revises in the next bar and the critic listens again before it signs.
Footnotes
Every note is a trace entry: who played, which tool, when it started and ended, its tokens, and its cost at a sample price of $0.003 per 1K tokens.
Budget
Each run carries a token cap. A guard checks every bar before it plays and halts the run rather than overspend; the console lets you lower the cap and watch it happen.

The architecture

Under the score, a scheduler, a guard and a trace.

One small engine plays every run. It is plain TypeScript, shared by this page and the console, and covered by tests.

passsend backONE BAR, IN PARALLELGoalin your wordsPlannerwrites the partsSchedulergroups into bars1 to 3 at a timeBudget guardhalts before overspendResearchweb_searchExtractstructured_outputScorerun_code, generateCriticlistens, signsFine.verified resultTrace: start, end, tokens, costLEDGERONE BAR, IN PARALLELsend backpassGoalin your wordsPlannerwrites the partsSchedulergroups into bars1 to 3 at a timeBudget guardhalts before overspendResearchweb_searchExtractstructured_outputScorerun_code, generateCriticlistens, signsFine.verified resultTrace of every stepLEDGER
  • Scheduler

    Groups the plan into bars. Steps in one bar start and end together, and no bar starts before the last one ends. Allow one worker at a time and the same goal plays in more bars.

  • Budget guard

    Adds up what the next bar would spend before it plays. If it would push the run past its token cap, the run halts there and the cut notes stay in pencil.

  • Trace and ledger

    Every step reports its start, end, tokens and cost as it happens. The console keeps each finished run in a ledger, with totals for every agent and tool in the session.

Live in your browser

  • The scheduler, parallel bars and timing
  • Send-backs, revisions and the critic’s sign-off
  • The budget guard and its token cap
  • Token and cost accounting, the trace and the ledger

Sampled for this offline demo

  • What each agent says, and each goal’s result
  • Token counts per step, fixed for each goal
  • A flat sample price of $0.003 per 1K tokens
  • Your own goal plays a generic five-step plan

The stack

Printed with a small, ordinary toolkit.

Colophon

Framework
Next.js 16 with the App Router, React 19 and TypeScript
Engine
A pure TypeScript simulation, lib/sim.ts, shared by the landing and the console
Engraving
SVG drawn from the run model, with a treble clef outlined from Noto Music
Type
Fraunces, Libre Franklin and IBM Plex Mono
Styles
Tailwind CSS 4 and hand-ruled CSS: hatching, stippling and line weight, no color
Checks
Vitest for the engine and the engraving, ESLint, TypeScript strict
Hosting
Netlify. No backend, no API key. Usage analytics by PostHog, with what you type masked

Liner notes

The case study on rikenpatel.work tells the story behind Conductor and how it was designed and built.

Read the case study

  1. IPlay goal IV and watch the critic send a step back.
  2. IIAllow one worker at a time and count the extra bars.
  3. IIILower the token cap to 1,500 and watch the guard halt the run.
  4. IVSilence the critic and see what ships unverified.

Take the podium.

Pick one of eight goals or compose your own, then watch the planner, the workers and the critic play it bar by bar, with every step kept in the ledger.