Reliability for agents in production
Find the change that broke your agent.
Corollary replays every failed run against the last good version and tells you which prompt, tool, or model swap caused it.
See what shipped this week→COR. 1.1ERROR PER PASS
COR. 1.2CONVERGENCE
COR. 1.3LATENCY, 240 RUNS
northwindHALCYON/QuillbankTESSERAorbitlyParcelworks
Failure taxonomy
Agents in production fail in unpredictable ways.
Most runs are fine. The ones that are not look fine too, until a customer notices. Scroll to sort a day of traffic by how it broke.
How it works
Surface failures you didn't know to look for.
Three views of the same run. Pick one.
ISS-41Refund tool called without order idRegression38 runs
ISS-40Retrieval returns archived policy docRegression12 runs
ISS-37Plan loops on ambiguous dateInvestigating7 runs
ISS-33Escalation skipped for VIP tierResolved21 runs
Introduced in v41, when order_id became optional in the refund tool schema.
CorollaryAPP 9:42 AM
New regression on support-agent after deploy v41.
Refund tool called without order id. 38 runs in the last hour, 0 in the previous 7 days.
Open issueReplay on v40
COR. 2.1PASS RATE BY BUILD
Ship with confidence.
Send us a work email. We'll set you up with your own traffic, not a sandbox.
Thanks. We'll reply to within one business day.