Today most agents run like this:
model → tool call → execution
Which means an LLM decision can directly trigger things like deleting data, transferring funds, deploying code, or sending emails.
CASA inserts a deterministic gate between the proposed action and execution.
The agent action is converted into a Canonical Action Vector and evaluated before anything runs. The gate returns one of three outcomes:
ACCEPT GOVERN REFUSE
Latency is ~53 ms on commodity hardware and the decision process is deterministic (no second model calls).
The repo has adapters for LangChain, OpenAI function calling, and CrewAI so existing agents can route actions through the gate.
Interesting early signal: the repo was cloned ~180 times by ~87 unique developers in the first ~60 hours after release.
Curious what people building agent systems think about the idea of a deterministic execution control plane.
cherndon222•2h ago
Today most agents run like this:
model → tool call → execution
Which means an LLM decision can directly trigger things like deleting data, transferring funds, deploying code, or sending emails.
CASA inserts a deterministic gate between the proposed action and execution.
The agent action is converted into a Canonical Action Vector and evaluated before anything runs. The gate returns one of three outcomes:
ACCEPT GOVERN REFUSE
Latency is ~53 ms on commodity hardware and the decision process is deterministic (no second model calls).
The repo has adapters for LangChain, OpenAI function calling, and CrewAI so existing agents can route actions through the gate.
Interesting early signal: the repo was cloned ~180 times by ~87 unique developers in the first ~60 hours after release.
Curious what people building agent systems think about the idea of a deterministic execution control plane.