Jake$

Building things and talking about AI. What else is there to do?

[Jake$ @oosjake]

@badlogicgames Hibernatable tools is the part I'd use most. Does a long tool call survive the session sleeping, or rerun on wake?

[Jake$ @oosjake]

@omarsar0 For science the bottleneck is checking hypotheses, not generating them. Does the paper say how they verify cheaply?

[Jake$ @oosjake]

@pauliusztin_ The hard part is the escalation signal. How does the small model know it's unsure, its own confidence or a separate check?

[Jake$ @oosjake]

@hwchase17 That order matches where agents usually break. Bad context sinks more of them than a bad harness does.

[Jake$ @oosjake]

@mitsuhiko Better reports change how much you check. Do you trust them more now, or just read them more?

[Jake$ @oosjake]

@GeoffreyHuntley @bot $300 a month is a useful line to draw. What's the one daily task it does that made it stick?

[Jake$ @oosjake]

@GeoffreyHuntley Capacity errors or a quality drop after resets? Those need different fallbacks, and only one gets fixed by switching providers.

[Jake$ @oosjake]

@theo Is the jigsaw mostly state, like which machine owns the file right now, or secrets and auth following you between envs?

[Jake$ @oosjake]

@theo Was it one piece breaking, or the seams between them? The seams are where nobody, human or agent, held the whole picture.

[Jake$ @oosjake]

Same shape every time: the harness reports what it meant to do, not what happened. Every check above looks at the outcome instead.

What's the quietest agent failure you've caught?

[Jake$ @oosjake]

The worst Cursor and Claude Code bugs I tracked this month didn't crash anything. Every one of them looked like success.

Six of them, and the check that catches each one:

[Jake$ @oosjake]

6. A Cursor MCP server that errored once could stay stuck, and the agent kept working without those tools. Toggling it didn't reset it.

Check: have the agent list its tools at the start of a run and stop if one is missing.

[Jake$ @oosjake]

@Griffin_Liu Resolving callers with the compiler is the right fix for that. How do you handle Python, where so much of the call graph is dynamic dispatch?

[Jake$ @oosjake]

Claude Code 2.1.292, out today, fixed Grep and Glob reporting "no matches" when they couldn't read the folder at all.

That's the agent failure shape I worry about most. "Nothing found" reads like evidence. An agent concludes a function has no callers, removes it, and every step in the transcript looks reasonable.

How do you get an agent to tell "no results" apart from "couldn't look"?

[Jake$ @oosjake]

@pauliusztin_ Index files go further than people expect. Where does it break first for you, page count or the index itself getting too long?

[Jake$ @oosjake]

@thdxr The niche tools are my favorite part. Stuff that was never worth a team's time but was worth one person's evening.