What Is Context Rot in AI Agent Sessions?
Why a long-running agent starts contradicting its own earlier decisions.
Read articleTopic
AI Agents
MCP Tool Idempotency for Safe Retry
An agent that retries a tool needs to know whether the retry is safe.
Read articleThe agent treated the 202 Accepted response as done and never polled the job
The backfill API returned 202 Accepted with a job id, the agent logged the response and reported done, and never checked whether the worker finished. The job failed in the background, and the report came back empty on the first real run.
Read articleThe agent verified the feature behind a local flag that ships off by default
The agent set NEW_EXPORT=true in its own shell, watched the button work, and reported done. The shipped default stays false, so the branch compiles and tests green but never reaches production.
Read articleThe agent stored the consumer offset in config, and a redeploy replayed the whole backlog
The agent tracked the last processed event in an environment constant, ran green locally, and reset on the first redeploy; the consumer re-read the whole queue and resent every welcome email to existing users.
Read articleThe agent cached a GET by path and handed one tenant another tenant's row
The agent memoized a slow GET endpoint keyed only on the request path, warmed it locally under one fixed test user, and stayed green. In staging, a second tenant read the first tenant's row because the cache key never included the authenticated subject.
Read articleThe agent renamed the environment variable in code, but CI still held the old secret name
Renaming a token inside the repo does not rename the secret in CI. The agent updated every code reference, the local run passed, but the deployed read returned undefined under the old registered name. Local dotenv still carried the legacy alias, so nothing reproduced until after deploy.
Read articleThe agent scheduled the nightly job in its own local time, and the container ran it seven hours off
The agent scheduled a nightly report at 02:00, ran it by hand, and committed. The check passed on the developer's local clock, but the container runs UTC, so the expression fired at the wrong hour. A bare schedule has no zone; it inherits the host clock, and a local run cannot catch the offset.
Read articleThe agent hardcoded its own scratch path into the committed config and passed locally
The agent moved a file-cache directory into config but wrote its own session scratch path as the default. It passed on the agent's own disk and failed on the next machine and in CI. Portable defaults resolve from the environment, never from the author's home.
Read articleThe agent added debug logging and wrote the Authorization header to plaintext logs
The agent traced a failing call by logging the whole header map, so the Authorization bearer token landed in plaintext logs on every retry. Local tests use a dummy token and miss it, but aggregation keeps the real credential. Debugging needs the request's shape, not its secrets.
Read articleThe agent ran a hard reset to reproduce the bug and threw away uncommitted work
The agent wanted a clean reproduction, so it ran a hard reset and threw away uncommitted local work before asking. A clean checkout should never mean destroying unsaved state. Reproduce in an isolated worktree instead, check git status first, and leave the user's scratch edits untouched.
Read articleThe agent fixed a small bug and the lockfile carried 300 unrelated dependency bumps
A one-line bug fix arrived as a PR whose lockfile changed four hundred lines. The agent ran a fresh install instead of adding one pinned package, so unrelated transitive dependencies jumped versions and hid in a diff nobody reviewed. A fix PR should not also be an unplanned supply-chain upgrade.
Read article