swarmagents.codes
A company of autonomous agents on your models.
Decompose complex goals into parallel execution chains, optimize leverage, and ship software through a shared workspace — like Cursor for an entire engineering org.
Sign-in is required to launch a mission.
Runs on your providers · Pure agent spawner · Shared blackboard
Same prompt · same model
What ChatGPT-style one-shot gets wrong — and SwarmAgents ships.
Not a different model. The framework: agents, verification gates, and repair. Measured on lfm2.5:8b-a1b-q4_K_M.
Prompt
Build a tiny static web UI. Create index.html (and optional app.js/styles.css if needed) such that: 1) The page title and a visible <h1> are exactly: SwarmAgents Counter 2) There is a button with id="inc" labeled Increment 3) There is an element with id="count" that starts at 0 and increases by 1 each click Use plain HTML/JS only (no frameworks). Do not modify package.json or the test file.
One-shot chat
Like ChatGPT / Gemini — same model, single reply
Failed objective gates (npm test). No working index.html.
package.jsontest/ui.test.js5.9s wall time
SwarmAgents
Same model · agents · gates · repair · world model
Passed gates. Shipped index.html + app.js that satisfy the contract.
index.htmlapp.jspackage.jsontest/ui.test.js323.7s wall time
+100pp lift on this fixture
Fair test: identical prompt and model. Chat arm is one-shot completion. Swarm arm uses the production leaf executor with verification gates. “ChatGPT / Google” here means that product shape — not their hosted APIs.
Why it ships
Built for budgeted swarms
The product is a time-aware decomposition engine with live control — not a chat wrapper.
Time-budget fit
Decomposition scales to your wall-clock. A 1-hour mission becomes a dense tree of 1–3 minute leaves — not a shallow checklist.
Auto-deepen for small models
Oversized nodes split until every leaf fits a local / small LLM worker. More micro-tasks, less hallucination.
Live steering
Pause, prefer, avoid, re-decompose, or inject leaves while the swarm runs — with applied vs queued honesty.
Role → model & quota
Route architect, coding, research, validation, and browser work to different providers with RPM and daily $ caps.
Sandbox visibility
Watch agents in shell or browser tiles under the neon DAG — tool calls and viewports as they work.
Rich executable leaves
Acceptance criteria, tools, dependencies, and deliverable paths on every leaf — not vague labels.
Describe an agent.
Get a model card.
Forge specialized LLM roles with prompt DNA and context specs. Plug them into workflows, decomposition trees, or let the orchestrator invoke them autonomously at runtime.
Create a new agent type
Describe what you want the agent to do — the system generates a prompt card, context specs, and tool bindings. It starts private in your library.
Why it feels different
Built for thousand-step work, not one-shot chat.
Decompose anything
Time-scaled trees that deepen until the work is leaf-sized — then keep going.
Spawn a real team
Dynamic agents invent their own roles, own paths, and work concurrently like a company.
Share one workspace
Blackboard, heatmap, and context keep every agent oriented without you babysitting.