01 / The brief
Keep intent intact while the factory changes around the agent.
The agent must sequence work, call tools and recover from interruptions without turning a local error into a full operational failure.
Illustrative case format / 01
Making a digital factory agent recoverable.
A controlled work environment for testing planning, perception and recovery when the state is incomplete, tools fail and constraints change mid-task.
02 / Pressure map
- Partial inventory and delayed sensor updates
- Tool timeout during a dependent action
- Conflicting work orders from different operators
- Blocked route that forces a plan revision
03 / What we build
A replayable world, not a one-off demo.
We model the task graph, actors, permissions, tools, state transitions and evaluation hooks so each run can be seeded, compared and repeated.
04 / What the team keeps
- Scenario catalog with severity and trigger conditions
- Full trajectory and tool-call traces
- Failure taxonomy mapped to remediation owners
- Regression suite for the next release
Make recovery
measurable.
This is the shape of a Broken Agents engagement: take the workflow you are not ready to trust, build the smallest world that can break it, and leave your team with evidence they can keep running after handoff.
Bring us your workflow
We will scope the first pressure map with you.