agent retried the same GitHub 401 for 34 minutes
left Cursor on a "just fix the CI" task while I made coffee.
came back to a wall of the same line: GET /repos/.../actions/runs 401. agent did not stop. did not ask. just kept retrying with the same expired PAT.
billing page said about $19 of sonnet for one stuck loop. I had no hard timeout on tool calls. that is on me.
now every agent run gets a max tool budget and a kill after 3 identical errors. messy english, but monday was quieter.
anyone else catching silent retries before finance does?
3 comments
Join the discussion
Log in to comment.
this is the Zapier class of bug, just with a higher invoice.
I once left a Make scenario on "retry on error" overnight. same 401 to Notion, 200 attempts, zero alerts. I only noticed because a board stayed empty.
rule now: if the error string is identical three times, stop. screenshots before Run still don't catch a loop that starts after you walk away.
same class of pain on ollama last week. agent kept calling a dead tool endpoint, wifi was fine, model was fine, the mcp was just asleep.
i now put a hard ceiling on identical errors before the loop can burn another euro. three strikes then kill. boring. cheaper than explaining the bill.
hit the same loop friday with a dead github fine-grained token. agent just kept doing
gh apilike the 401 was a temporary wifi blip.i wrapped tool calls so identical stderr lines count up. after 3 it dumps the last call to /tmp and exits 1. still burned ~$8 of sonnet before i noticed — finance already pinged.
do you kill on status code, or exact string match?