Claimswarm

A Mozaik Hackathon 2026 entry: a concurrent claim-investigation swarm, not a sequential agent pipeline.

Submit-form paste: CONCURRENCY.md is the written answer to “How do the agents run concurrently?” Official rules (updated 4 Sep 2026): https://build.jigjoy.ai/rules — submit at https://build.jigjoy.ai/submit

You state a claim. Steelman, Redteam, and Contextualist join() one runtime and start runLoop at the same time. Sentry reacts to function_call.started while those loops are still running. A policy InterceptionHandler aborts overconfident settlements mid-loop. Synthesizer updates a live board as function_call.completed arrives. The operator can inject a message mid-run. When Scout leaves, Auditor join()s — reassignment is a situation handler, not a controller.

Built from the official cli-agent-starter on @mozaik-ai/core ^4.0.5 (current on npm; the older ^3.13 / AgenticEnvironment names are not what this package exports).

Organizers: JigJoy × daily.dev × Hyperskill. Docs: https://docs.jigjoy.ai/

How the agents run concurrently

This section is what the submit form asks for. The paste-ready version is CONCURRENCY.md.

Claimswarm is a Mozaik v4 swarm (defineRuntime + join() + fire-and-forget runLoop), not a sequential pipeline.

Three specialist agents — Steelman, Redteam, and Contextualist — each join() one runtime and keep their own ModelContext. When the operator sendMessages a claim, every specialist’s message.sent handler calls runLoop and returns immediately. Mozaik does not await runLoop or processor.apply, so the three loops think, call cite_evidence, and answer at the same time. npm run demo prints ISO timestamps and exits 0 only if the third inference.started occurs before the first of those inferences completes.

They share the environment, not memory. Shared data lives on ClaimswarmState (claim, evidence board, sentry flags, roster) plus the event bus. Coordination is reaction:

• Sentry is a joined policy participant. On another agent’s function_call.started it flags the call while that caller is still inside its loop.
• Each specialist’s runLoop carries an InterceptionHandler. If a model_message tries to declare the claim “settled” / “proven beyond doubt”, the handler rewrites the transition to message_received mid-loop (interception.started / interception.finished) and that agent infers again. Other specialists are not paused.
• Synthesizer listens to function_call.started, function_call.completed, and model.answer and updates the live board as events arrive. It does not wait for a join barrier.
• The operator can sendMessage again mid-run; new loops start while earlier ones are still open.
• Coordinator watches participant.left. When Scout leaves, Auditor join()s the same runtime and starts investigating.

Nothing is one agent with tools in a for-loop. A join-and-wait fan-out cannot produce the overlap proof the demo requires.

flowchart LR
    Operator -->|"sendMessage claim"| Runtime
    Runtime -->|"message.sent"| Steelman
    Runtime -->|"message.sent"| Redteam
    Runtime -->|"message.sent"| Contextualist
    Steelman -->|"runLoop fire-and-forget"| Runtime
    Redteam -->|"runLoop fire-and-forget"| Runtime
    Contextualist -->|"runLoop fire-and-forget"| Runtime
    Runtime -->|"function_call.started"| Sentry
    Runtime -->|"function_call.started"| Synthesizer
    Runtime -->|"function_call.completed"| Synthesizer
    Sentry -->|"InterceptionHandler rewrites model_message"| Steelman
    Scout -->|"participant.left"| Coordinator
    Coordinator -->|"join Auditor"| Runtime
    Operator -->|"mid-run sendMessage"| Runtime

Loading

Environment is shared (ClaimswarmState + the event bus). Each agent keeps its own ModelContext. Nothing is a single agent with tools in a for-loop.

How to run

npm install

2–4 minute demo path (no API key)

npm test
npm run demo

npm run demo is the judge path. It uses SimulatedInferenceRunner so runLoop still runs — it just does not call OpenAI. Watch for:

1. Three inference.started lines whose timestamps overlap.
2. function_call.started then sentry.reacted before that caller’s loop goes idle.
3. board.updated on function_call.completed (Synthesizer, not a final join).
4. interception.started when Steelman’s overconfident settlement is rewritten.
5. A mid-run operator inject, then Scout leaving and Auditor joining.

The process exits 0 only if overlap + sentry + interceptor are proven.

Interactive ops room

npm start -- --sim

Type a claim, press Enter, then type another line while the board is moving. /leave drops Scout (Auditor joins). /exit quits.

Live OpenAI

cp .env.example .env   # set OPENAI_API_KEY
npm start -- --live

Same participants, same interceptor, same tools. The model is gpt-5.4-mini. Without a key, the TUI falls back to dry-run.

Architecture

Participant	Role	What it reacts to
Operator	Human	Sends the claim and mid-run injects
Steelman / Redteam / Contextualist	Agents	message.sent from others → runLoop + InterceptionHandler
Sentry	Policy observer	function_call.started, interception.*
Synthesizer	Live board	function_call.started/completed, model.answer
Coordinator	Roster	participant.joined / participant.left → Auditor join()
EventLog	Transcript	Loop events with ISO timestamps
Scout / Auditor	Reassignment	Scout leaves; Auditor joins mid-run and starts investigating

Tools (cite_evidence, publish_verdict) search a local public-finding corpus (Bloom/Ctrip 2015, Gibbs et al., Microsoft Work Trend Index, Barrero/Bloom/Davis). No search API. No invented papers.

Tests

npm test drives the real runtime:

• participants join() and keep separate contexts
• a simulated function_call.started makes Sentry (and Synthesizer) react immediately
• function_call.completed updates the board
• InterceptionHandler rewrites an overconfident model_message
• three runLoops overlap in time; Scout leave assigns Auditor

Contest notes

Official rules (last updated 4 September 2026): https://build.jigjoy.ai/rules

• Must use @mozaik-ai/core. This repo depends on ^4.0.5 and builds with defineRuntime, createAgent / createHuman, runLoop, InterceptionHandler, and structured ContextItems. Current docs: https://docs.jigjoy.ai/
• Two or more concurrent agents. Three specialists run overlapping runLoops; Sentry is a fourth live participant. A single-agent project does not qualify.
• README concurrency write-up. See the section above and CONCURRENCY.md for the submit-form paste.
• Deadline Monday 7 September 2026 09:00 CET. Resubmits allowed; latest entry is judged.
• Starter/boilerplate allowed; this entry is the swarm substance (not the one-agent Ink template). AI-assisted development is allowed.
• Solo entry. Teams may be up to 4; this is one person.
• Judging weight: genuine concurrency, execution, creativity, demo clarity. npm run demo is the 2–4 minute path; a screen recording of it is the intended demo video.
• Cash prizes are paid through Upwork per the official rules. No payment details are stored here.
• MIT license. No secrets in the repo. OPENAI_API_KEY is optional and read from the environment.
• Dry-run is not a fake test: it still constructs a Mozaik runtime, joins participants, and runs runLoop.