Field Notes
What broke → what we changed
Short, honest notes from autonomous engineering runs
Real failures, root causes, and the concrete process changes that followed across our autonomous engineering projects.
PlanningFinance project
The limit no portfolio could meet
A spec risk limit that the strategy's own design made impossible to meet was caught mid-build when the system safely refused to trade.
✕Before(what broke)
The system safely refused to trade mid-build after encountering a risk limit that the strategy design could never satisfy.
✓After(what we changed)
Plans now raise strategy questions to the human before any code is written.
PlanningRisk limitsStrategy
ReviewData pipeline
Four rounds to find three bugs
A diff-only review found one issue per round. Replaying the change through the deep review found all of them in one pass.
✕Before(what broke)
A diff-only review found one issue per round across successive review cycles.
✓After(what we changed)
Read-only checkout plus an acceptance checklist caught all issues in one deep review pass.
ReviewData pipelineChecklists
ConfigFinance project
Defaults that quietly drifted
Shipped safety thresholds drifted from the written spec for several sprints, while every test passed.
✕Before(what broke)
Safety thresholds drifted from the written spec across multiple sprints while every test continued to pass.
✓After(what we changed)
Spec values are pinned by tests, and reviewers diff spec against config.
ConfigTestingSafety defaults
ToolingInternal harness
The agent working in the wrong folder
An agent kept finishing tickets with nothing to commit. The cause was the launcher's working directory, not the model.
✕Before(what broke)
An agent repeatedly finished tickets with nothing to commit due to an incorrect working directory in the launcher.
✓After(what we changed)
Fixed the launcher and verified working directory state with a live probe.
ToolingHarnessVerification