A multi-agent run that has to prove the result

The audit checks the contract. The MECHA run executes it. Your compiled work order gets dispatched to a real multi-agent swarm inside an isolated Daytona sandbox, and the run is not done until the acceptance criteria are verified by a separate reviewer.

How a run works

1. You build a work order. Goal, acceptance criteria, non-goals, budget. Grok can draft it from plain English.
2. Pick a strategy. The swarm fans out under senate, triumvirate, best-of-3, or a solo lineage, with workers from Claude, Codex, and Grok lines via OpenRouter.
3. Watch it run. Live telemetry streams from the sandbox as workers post evidence. Runs take minutes, not seconds.
4. Get the verdict. A reviewer synthesizes the final answer and the full evidence chain. You get the report, the exit code, and the artifacts.

Pricing

Runs start at $10. Scale the number of agents and the price moves up with the compute you are actually using, capped so a runaway swarm cannot surprise you.

Why sandboxed matters

The sandbox keeps the swarm from touching anything outside the job. No stray pushes, no emails to real customers, no side effects beyond the workspace you gave it. Combined with the contract's explicit preauthorized list, that is the difference between an agent that explores and an agent that wanders.

When to run MECHA

• When the job is real and the contract has been audited
• When you want the result verified by an independent reviewer, not the same model that did the work
• When a single chat answer is not enough evidence for a decision

Run the $1 audit first, then build the work order that survives it.