Run one agent task. Check the result.
Follow three guides on scheduling, completion checks, and uncertain writes. Start with one task whose output you can inspect.
More recent articles
-
Claude Opus 5.5 vs. Sonnet 5.5: which model fits your work?
Compare Claude Opus 5.5 and Sonnet 5.5 using official benchmarks and API prices, then choose a model and effort level for coding, documents, research, and agents.
-
A delegation wrapper earns its keep by refusing flag combinations
A CLI validates flags one at a time; nothing validates the combination. After a read-only run wrote to disk because a second flag outranked the first, the wrapper…
-
Qualify an agent CLI for unattended use in four probes
Before an agent CLI runs on a schedule, four cheap experiments establish what its flags actually do: whether restrictions hold, what refusal looks like from the outside, how…
-
Reasoning effort changed what my agent tried, not how well it did
Twelve timed runs of one arithmetic task across four effort levels. Latency did not separate the levels and neither did correctness. What separated them was behavior: the two…
-
A five-minute cache made my fixes look like failures
An audit kept reporting seven broken links that no live page contained, because its input list came from a cache with a five-minute lifetime. The same edge also…
-
Keep your agent’s receipts where the cleanup cannot reach them
Every receipt and pre-change backup from a three-day operation lived in a scratch directory under /tmp, and a routine cleanup erased all of it. What survived was the…
-
Your scheduler cannot promise the job ran, so let the job decide
launchd documents that interval firings are dropped when the machine sleeps and coalesced when calendar intervals pile up. Neither mode can tell you the work happened, so the…