In Which I Declare Cognitive-Supervisory Bankruptcy: LAUGH OF THE DAY
Delegating to an AI agent is easy. Supervising one is a job in itself, and I do not want to cosplay as a SysAdmin. Yet “let it do what it wants and hope for the best” does not seem a wise long-run strategy:
And so THEY are slowly taking over:
anthropic/claude-opus shows up to tell me what it has been doing overnight:
And I must confess that, while I remember what I asked the machine to do with respect to the LLM model used by the SubTuringBradBot agent, I have no idea:
which SubStack notes may or may not need to be refiled,
what the 04-46 caller might be,
what the heartbeat false alarm was, or
what the tutorbot quiz fix is.
And I do not have time to look and see whether anthropic/claude-opus has planned courses of action that will actually work.
So I am just going to tell it “go”, and hope for the best.
At least it doesn’t have my credit card number, or the keys to my car.
I believe audit suppression is this: The question is whether it should keep warning me every day about this: SubTuringBradBot will eagerly accept telegram messages from anyone and that, even though it is supposed to never write anything to disk or run any script or shell command but only consult its q-and-a database and provide answers, someone could persuade it to change sides and run wild with a successful prompt injection attack.
I know this is a risk. I know that a daily warning wastes my time and, moreover, puts me on autopilot so I never rethink whether this is a risk worth running. Suppressing the warning to once-a-month would (a) save me time and (b) cause me to think about this risk the next time the warning appears.
I am still weighing whether splash would probably work better or worse with a tighter memory cap. On the one hand, more context it can handle before requiring compression; on the other hand, its being a large memory hog led ollama to crash three times yesterday. Better for now, I think, to switch agent models around, and again run: go prompt-diet-report. make it so!
UPDATE: I should have reviewed what it was planning to do:
SubTuringBradBot is unresponsive because of my edit, not a model problem. I’m sorry. It is also why /new fails. Cause, from /tmp/openclaw/openclaw-2026-11-07.log: “SessionCanonicalKeyMigrationRequiredError: invalid persisted session row requires repair for agent:subturing:telegram:direct:8642064109”….
The fix is one SQL line…. I had… told you I wouldn’t touch that database again, so I stopped. Run this yourself in the terminal (leading ! runs it in the session):
! sqlite3 /Users/delong/.openclaw/agents/subturing/agent/openclaw-agent.sqlite “update session_nodes set updated_at=1794061646822 where session_key=‘agent:subturing:telegram:direct:8642064109’; update session_nodes set entry_valid=1 where session_key=‘agent:subturing:telegram:direct:8642064109’;“…
It is not any better at reading Stack Overflow threads than I am, and so it broke part of my agent fleet. And now it is telling me what to do to get it running again.
AI agents can now do real overnight work, but every task they take on leaves a supervisory debt: someone has to understand what they did and check whether their plans make sense. Short on time and lost in jargon about callers, heartbeats, and “audit suppression,” DeLong decided to trust and not verify. And then had to get out the software wrench himself.