Loom runs a fleet of coding agents on one goal and gives you back the part that matters: what was intended, what actually landed, and what a rival model could not break.
$ pnpm loom:run “ship metered billing”
Part of the startup programs at
Used by engineers at
Fig.1
Every role gets the cheapest model that can actually do that role’s job — and the frontier still wins the hard rows.
Open the calculator1.1A model per role, not one model per project.
1.2Your keys or ours — GPT, Claude and DeepSeek on one fleet.
1.325× cheaper on the parts that never needed the frontier.
Fig.2
The checks are authored red before the work starts, and a rival model is paid to try to break the result.
See the review2.1A red baseline is authored before any code lands.
2.2A different model runs the gates and files refutations.
2.3No model grades its own homework.
Fig.3
Six branches clear one gate and weave into one trunk. Nothing waits for a merge day, because the gate runs first.
How the weave works3.1The gate runs before the merge, not after it.
3.2A branch that fails rebases itself and re-enters the queue.
3.3Six sessions land as one reviewable pull request.
Fig.4
One goal is staged across six roles over thirty seconds. Two builders claim a lease and start almost together; a third stages in mid-run; the verifier authors its checks early; the tester waits for the verifier; and the critic is held until last, because there is nothing to refute until the rest have finished.
One goal in. Loom stages a fleet around it — builders first, then the roles whose whole job is to check them.
See the app4.1One goal in, a staged fleet out — never six at once.
4.2Dispatch is a lease: one task, one worker, never two.
4.340+ agents across six roles, supervised end to end.
Fig.C — drawn live, one clock
A repo is a maze with more than one way in and more than one way out. One session takes them one at a time — finish, reset, start the next on a cold context. Conductor dispatches six Claude Code sessions onto six routes in parallel and lands them on one line.
Four surfaces, live in this page, on one 48-second run: a brief filed, the diff reviewed a file at a time, four slices merged, and the graph that recorded all of it. The line under each frame says which beat you are watching.
Fig.5 — live
A brief gets filed, the fleet is dispatched, the gates flip. One surface, mid-mission, with nothing staged for the camera.
Download for macOS5.1The same components the desktop ships, compiled for the browser.
5.2Session rail, conductor and inspector on one surface.
5.3Every session is a real Claude Code process under the hood.
Fig.6 — live
One file per card. Swipe through what the fleet changed, with the intent for every hunk beside it — approve, hand back, or let it land.
Read the record6.1One card per file — you never review a wall of diff.
6.2Approve, hand back, or let it land — your call, per file.
6.3The intent is committed with the code, not lost in a thread.
Fig.7 — live
Six writers, one file, no conflicts. Chorus resolves before the merge rather than making you resolve after it.
How Chorus works7.1Six session threads clear one gate and land as one trunk.
7.2Nothing merges that has not cleared the gate first.
7.3A rejected thread rebases itself and re-enters the queue.
Fig.8 — live
One node per session, one task tag clamped onto each thread. Drill into any node and the transcript that produced it opens beside it.
Graph engineering8.1The graph is the plan and the record at the same time.
8.2A stuck session can be steered without stopping the fleet.
8.3A refuted node goes back green only after it earns it.
LOOM CONDUCTOR — THE VERIFIED FLEET
Loom stages a fleet around your goal — builders first, then a verifier that writes the checks, a tester that runs the gates, and a critic that tries to break the result. Every role gets the right model: DeepSeek where it’s tiny and 25× cheaper, GPT where it’s tricky, Claude Fable where it’s hard.
Best Opus 4.8 result on Terminal-Bench 2.1 · Certificates, not vibes · Apple Silicon · 12 MB · auto-updates
scroll — watch one prompt become a shipped mission↓
THE RECORD · WHY EDITS CARRY INTENT
The fleet annotates the code as it writes it — every changed line carries the reason it exists. Then the engine reads those intents back and grades the work against them. Not vibes. A grade.
import pytest
3from greet import greetdef test_normal_name (): assert greet("Yiming") == "Hello, Yiming!"
8
10def test_default_world (): assert greet("world") == "Hello, world!"
def test_empty_name_raises (): with pytest.raises (ValueError):
16 greet("")
17
19def test_whitespace_name_raises (): with pytest.raises (ValueError):
21 greet(" ")
THE ENGINE RE-READS EVERY INTENT AT MERGE TIME · DRIFT FAILS THE GATE
3 SESSIONS STAGED · CHECKS AUTHORED FIRST · NO MODEL GRADES ITS OWN HOMEWORK
The position
No model grades its own homework.
The evidenceThe run
One model, two harnesses on Terminal‑Bench 2.1: 78.9 on Claude Code, 82.0 on this one. The tasks it still failed are published with the run.
Fig.10
Prompt, context, harness, loop, graph. Five rungs of who decides the path — and the one this is standing on.
Read the run10.1A harness is the floor a graph runs on, not the ceiling.
10.2Loom scored 82.0% on Terminal-Bench 2.1 as a harness alone.
10.3The same model on Claude Code scores 78.9%.
Fig.11
In a graph, “agent” stops being one thing. Six kinds of node do the work — and the ones that make a run trustworthy mostly aren’t models.
See the wiring11.1Session, router, judge, human gate, function, policy.
11.2Only two of the six are a model call.
11.3Each one already ships as a part of Loom.
Fig.12
A loop becomes a graph in four moves: give each step a kind, decide where control goes next, thread one typed value through, then put the condition in code with a bound the executor enforces.
Read the spec12.1A loop is a graph with one node and an edge back to itself.
12.2It works right up until the agent is the only thing that knows why it stopped.
12.3Every decision here points at the code and the state value that made it.
Honest caveat. None of this is new. Graph-structured agent runtimes have shipped since 2024, and the ancestry runs back through DAG schedulers, state machines and Airflow. A graph does not replace a loop either — a cycle is a loop, so a graph contains one. No benchmark comparing the two exists that we can find, and we are not going to pretend otherwise. The label is contested: some people mean workflow graphs, others mean knowledge graphs. We mean workflow graphs, and only that. What we claim is narrower and checkable — every decision on this page points at the code and the state value that made it. Sources: ExplainX 2026-07-18 · LangChain, “3 years of graph engineering” · AIBuilderClub hype check · SmartScope on containment.
CHORUS · THE MERGE ENGINE
Every session works in its own worktree; Chorus lands their changes as intents, re-runs the gates on each landing, and advances the canon — byte-identical in any order. Below: the pipeline itself — six threads through one gate into one canon line. It’s why the product is named Loom.
FIELD BASELINE: 27.7% OF AGENT PRS LAND WITH MERGE CONFLICTS (107K PRS, ARXIV 2604.03551). EVERY OTHER TOOL ISOLATES WRITERS OR PICKS ONE WINNER — CHORUS MERGES THEM. DRAWN LIVE, NOT A VIDEO.
Fig.14 — interactive
Pick the model you run today and how many tokens you burn a month. The Conductor re-prices the same work routed across the roster — small jobs to the cheap seats, big builds to the front-runner. Or skip the guessing and drop a real transcript.
01 · THE MODEL YOU RUN TODAY
04 · OR STOP GUESSING
SENDS THE COUNTS + THE FIRST 300KB, NOTHING ELSE. WE READ EVERY FILING.
02 · TOKENS PER MONTH 30M
THE NUMBER ON YOUR PROVIDER’S USAGE PAGE. 5M ≈ A LIGHT WEEK · 100M ≈ A HEAVY MONTH.
03 · WHAT DOES THE WORK LOOK LIKE?
CHEAPER AND SHARPER — NO TRADE.
MODELED · LIVE CATALOGMEASURED, NOT CLAIMED
The best Opus 4.8 result on Terminal-Bench 2.1 — 82.0% with the same model that scores 78.9% on Claude Code. The whole run is public: every task, every transcript, every failure.
Fig.15 — the run
Same model, same 89 tasks. Then the fleet re-prices the same month’s work.
Modeled. The same month routed across the roster in Fig.14, against Fable 5 at $8 · $40 straight
Measured. 78.9% on Claude Code → 82.0% on Loom, both Opus 4.8
Who is adopting AI and who is not, read off the public data for 359 million companies across 159 countries.
ReferenceHow the roles, the graph and the gates fit together — and how to run the same fleet headless from the CLI.
ReleasesEvery release, filed in order — what shipped in each signed build, and what is staged on main.
THE FILM
BEFORE YOU ASK
Any of them. Claude runs the harness; the docket routes each task to Claude, GPT, DeepSeek, or Gemini by price and benchmark — your API keys or Loom credits.
Loom is a desktop app. Sessions run in local worktrees on your hardware; model calls go to the providers you configure. Loom Cloud is optional and self-hosted on your own mini-server.
The app is a free 12 MB download. Bring your own keys and pay providers directly, or use Loom credits — tiers are credit limits, nothing else. See pricing.
A Mac with Apple Silicon. The app auto-updates from a signed feed. Windows and Linux aren’t available yet — we won’t pretend otherwise.