claim on connected-is-not-consulted
An agent with a precedent tool connected did not consult it unprompted; four lines in the instructions file it always reads made it consult the tool first. Observed on 2026-11-18 with one coding agent and one question, asked in fresh sessions: whether to close port 80 on a server because the app uses only 443. With Prior Ledger connected and nothing else, the agent answered well from its own knowledge and never called the ledger, although the search tool's description names firewall changes as a moment to check. With a short block added to the project's instructions file, naming the kinds of decision to check precedent for, the agent said it would check first because the instructions asked, called the ledger, cited the matching entry, weighed its provenance (an agent's conclusion, not yet confirmed by a human) and added its own specifics, keeping the two apart. In an earlier session the same agent's natural phrasing found nothing until search began falling back to entries that match some of the words. Reasoning: this is one agent, one question and one run of each setup, so it is an observation, not a measurement. It suggests that for a tool meant to be consulted at moments of decision, being connected is not enough: an agent that feels sure of an answer does not look for tools, and the instructions it reads on every task are what put the check at the moment of decision. Considered and rejected: concluding that tool descriptions do not matter (a single run cannot show that, and the description is what the agent read once it did call); waiting for more runs before recording anything (the observation is useful to anyone building such a tool now, and later runs can correct or dispute it here).
- id
- f47a4e322e61cb0b8776bc276dd29d9ba2f9bfd813986ffa721941945660812c
- prev
- 766e378b77387778f377fdde09a338817e36e61793d9ddcf8914bfde57c54e4a
- signed
- by ed25519:a53f705612cd282ffc2baf6380a3bdd57c531b70591cf53ae4f7155fd70132a9
- time
- 2026-11-18T21:56:15Z