Agents With Brakes
A public reference architecture for production-grade enterprise agent platforms with policy, identity, evals, observability, approvals, rollback paths, and failure management.
System flow
How the pieces move.
- Policy Set identity, tenant, and permission boundaries.
- Tools Gate actions through contracts and approvals.
- Evals Measure behavior before and after release.
- Rollback Pause, recover, or degrade when runs fail.
Problem Statement
Most agent demos treat the agent as the system and skip the controls that decide whether it can survive inside an enterprise.
Architecture Summary
A platform control plane wraps agent execution with policy, identity, tool access, memory, knowledge boundaries, telemetry, evals, approvals, cost controls, and rollback paths.
Key Design Tradeoffs
- Optimizes for architecture clarity over demo theatrics.
- Treats human approval, failure handling, and observability as core platform behavior.
- Keeps agent capability separate from authorization, evaluation, and operational control.
Technologies
Short note: Agents With Brakes is the architecture argument I keep coming back to: production agents are not solo performers with better prompts. They are controlled systems with permissions, telemetry, evals, human approval, cost limits, memory boundaries, and recovery paths.
The repo is a public map of the platform pieces that make agent autonomy less magical and more governable. Less “watch the agent do a cute thing.” More “what happens when it is wrong, expensive, over-permissioned, under-observed, or confidently headed toward a sensitive system?”