Developers Urged To Guide AI Agents With Reasoning Frameworks Over Rules

Noooo you just destroyed my dream of being generic on prompts and context just telling codex to do it relying heavily on the auto compaction that just works. I mean a a few weeks ago was just go long let auto compaction do its job, let codex do its magic ... what am i missing here?

Build agents that only manage context. I call it the librarian, one of my councils. Those agents are each responsible for a subject area. They have layers.

However they essentially act as the fetchers of meaningful accurate context to provide the worker agents with the context they need, either proactively or on demand.

Don’t expect the worker to know/ discover everything themselves.

Do t expect the worker to know exactly where to find what information when they’re effectively a new start…

@canmachinethink You can’t rely on deterministic context stuffing over very long work with many autocompactions

@_lopopolo Can't we put stuff we need to be in context into the system prompt? Or is that unreliable as well after compaction? If yes there should be a post compaction hook where you can reprovide the important context.

@_lopopolo The "paved workflow" framing nails it. Rigid rule lists rot the second autocompaction blows context away anyway. What's worked for me: strong defaults + clear escape hatches, then let the agent reason about when to deviate. Curate the path, don't script every step.

@JohnThilen Seems good for human legibility, not sure what else it gives the agents

@_lopopolo I think @garrytan’s latent vs deterministic framework is really good for this!

@vadimcomanescu All of these things are the same lol

@_lopopolo So what's your take on extensive obsidian/notion/etc backed knowledge bases for agents? Sounds like they are at least over-engineered.

@_lopopolo that why we have the agents.md right? Also is context window length really a problem we will still have in the future?

are there any downsides to storing this all in the repo vs. trying to create a mechanism that pulls in reasoning docs JIT? I'm curious if I'm wrongly thinking it would give the agent less noise to sift through or get confused by? I'm picturing "tell the agent to seek within the repo" vs. "point the agent to a very specific thing outside the repo" with the thought process it then will not grep/read anything unrelated into the context window

@mitschabaude You don’t need to deterministically context stuff. Let the model decide

@_lopopolo Agreed. Rules collapse once the task graph gets messy. What holds up better is a harness that narrows action space, exposes context on demand, and makes recovery cheap.

@_lopopolo "For things to remain the same, everything must change." Tomasi di Lampedusa 😎

@_lopopolo This feels like the part everyone skips over. My guess is the winners won’t be the ones with the smartest agents, but the ones with the best systems around them.

@_lopopolo It’s pretty interesting when you consider specifically what must be trad functions vs reasoning when to call said function and reason on the output. Theres a lot things people do in traditional software to get to a great outcome that can help with agent design when mapped out

@_lopopolo /goal and workflows end up pretty great.

I really feel like this is the next shift. Workflows let you benefit from the stochasticism of the agents while still retaining control via a deterministic process. It makes such a difference in terms of reliability, output quality, and confidence. Things still go wrong and drift which is why I've found it incredibly important to also be able to steer mid run and interrupt agents inside of workflows as well to be able to course correct when things go wrong.

@_lopopolo @JohnThilen gives them negtok/s