Verified Agent Runtime
A team of AI agents that build software together.
Every handoff verified. Every artifact checked before the next agent touches it. Not trust — proof.
Latest run
Full team run: coordinator decomposed goal into 8 tasks, executed concurrently (max 3), verified every handoff between agents.
Five specialized agents
Each agent has a defined role, trust level, and verification requirements. No agent trusts another — every handoff passes through a verification gate.
Coordinator
Decomposes goals into dependency graphs. Schedules concurrent execution. Routes work to specialists.
Code Agent
Writes implementation code. Follows type-first contracts. Produces artifacts that must pass verification.
Review Agent
Inspects code for correctness, security, and edge cases. Rejects artifacts that don't meet standards.
Test Agent
Writes and runs tests. Verifies behavior matches specification. Catches regressions.
Research Agent
Gathers context from codebases, docs, and APIs. Provides evidence for implementation decisions.
Boundary verification at every handoff
When Agent A produces an artifact for Agent B, it passes through a verification gate. The gate extracts claims from the artifact, verifies each one, and only forwards verified work. Rejected artifacts return to the source with feedback.
How it works
Five stages from goal to verified output. Concurrent execution with dependency-aware scheduling. Retry with feedback on rejection.
Decompose
Coordinator breaks the goal into discrete tasks with explicit dependencies.
Schedule
Dependency graph determines execution order. Independent tasks run concurrently (max 3).
Execute
Specialized agents work on assigned tasks. Each produces typed artifacts.
Verify
Boundary verification at every handoff. Claims extracted, each verified independently.
Retry or deliver
Rejected artifacts return to source with specific feedback. Verified artifacts move forward.
Safety by design
Multi-agent systems introduce new failure modes. The runtime has built-in defenses for each one.
Collusion detection
Detects when agents consistently agree without evidence. Flags rubber-stamp patterns.
Kill switch
Strict hierarchy. Any level can halt agents below it. Immediate shutdown on safety violations.
Circuit breakers
Auto-halt on error rate thresholds. Prevents cascading failures across the agent team.
Trust dynamics
Rolling window trust scores. Agents earn trust through verified outputs. Trust decays on failures.
4-level immutability guardian
Core verification rules are immutable. Agents cannot modify the rules that verify them.
Rollback manager
Auto-triggered on error rate spikes. Reverts to last known-good state. No manual intervention.
Run it locally
The runtime is available in the CLI. Give it a goal. It decomposes, delegates, verifies, and delivers.
npx tryassay runtime fullteam "your goal here" --cwd /path/to/projectRuntime access is available on Enterprise plans and as a standalone product.