Resources / Blog
Notes from the engine room.
Why deterministic execution beats re-deriving the same workflow, and what we learned building the engine that does it.
- 7 min
A long policy is not a runtime
HANDBOOK.md finds that the best evaluated agent configuration completed only 36.2% of long, policy-governed processes without missing a requirement. The result points toward a different runtime model: let agents design the procedure, then execute the procedure as code.
researchreliabilitydeterminismread → - 6 min
Cori vs Agent Harnesses
Agent harnesses can reach the same outcome through different paths, with wide variation in execution time, tool calls, and tokens. Cori turns the settled path into code.
engineeringbenchmarksread → - 4 min
Scheduled Agent work does not scale
Your agent solves a problem once by searching. Then it searches again, identically, on every call after. The fix isn't a better agent — it's writing down what the agent found.
engineeringdeterminismread →