I've been thinking about why verifying AI agent output feels so much harder than writing the spec that produced it. That question led me to rethink where my attention actually belongs in the process, and eventually to build https://t.co/TpxaTqqqag https://t.co/UBbhJaFkFH
Every time an agent explores your code base it builds up its own mental model of how the code works, and then throws that model away when the session ends. Code walkthroughs turn that model it into a step-by-step guide to bring you up to speed on any part of the system.
Functions are the biggest cause of program failures
But seriously, subagents are just function calls. If you don't pass it the right arguments, how can it give you the right results?
https://t.co/XBxiYRz3Hs
Subagents are the biggest cause of AI failures
In most harnesses:
1. Main agent writes a summary of what it knows
2. Subagent starts from that summary only, and makes assumptions
3. Subagent works and sends back a new summary of what it did
Same flaw as compaction: all summaries are lossy. This is how your assistant makes emails you never wanted, or purchases what you told it just to get the price of
So I tried something new: checkpoint agents
The initial agent has complete context of what needs to be done. And that input is already cached
Fork the session where it is. Send the fork ahead to do the task and it will do exactly what the main agent would have done; their context is identical
What is even better is when the fork finishes, the cache of the main agent is kept warm by the subagent (they share the same prefix)
I measured it on a 137k-token session with Opus 5.5:
→ Main agent does it on its own: 67 s, $0.38, context +8-17k
→ Checkpoint: 117 s, $0.68, context +2k
→ Classic subagent: 123 s, $0.57, context +4-5k
But here is what I found: avoiding large input is even more important than keeping cache. And most subagent tasks were done better either in series with retrieval, or split up to be parallel over fresh sessions from the start
I do not think there is a place for subagent delegation at all. There should either be a super-agent that orchestrates and does no work, or a long running agent that can retrieve from previous sessions, not a worker that completes tasks partially and delegates the rest to other workers
Don't design harnesses based on our assumptions of human behavior
I modeled this after emacs (my favorite program of all time). The UI / parts of the sequencer are written in (a custom) LISP and can be hacked and extended without restarting the app.
More DAGs should be drawn this way. In a pipeline like the one below, the nodes form a list, so each node can be a table row with metadata beside it.
Parallel branches are the main limit, but horizontal space is cheap!
@JungleSilicon Yeah, and things like Mars rovers which have a bunch of custom softwy and hardware for specific tasks and go through a lot of verification