How we reduced state drift in multi-step AI agents (practical approach)

<!-- SC_OFF --><div class="md"><p>we run an agent thing in production and we use langfuse for traces.</p> <p>last month our agent started refusing requests it should have

<!-- SC_OFF --><div class="md"><p>Every time one of our agents made a bad call in production, we had the same problem. Logs showed what happened. Nothing showed why it decided to

<!-- SC_OFF --><div class="md"><p>I’ve been experimenting with a different way of handling agents and I’m trying to sanity check if this makes sense.</p> <p>Right now most setups

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1t5lywi/i_analyzed_922_agentic_task_trace_and_found_the/"> <img alt="I analyzed 922 agentic task trace and

<p>Everyone builds demo agents. They work great in the terminal. You show them to your coworker, they say "wow." You feel like a genius for about forty-five minutes.</p> <p>Then

<p>`---<br /> title: "Why I Built a Brain for My AI Agents — and What It Taught Me About Memory"<br /> published: true<br /> description: "Anamnesis v0.3.0: The cognitive

<p>Claude's context window holds 200,000 tokens. Gemini's stretches to two million. But response quality starts degrading long before the window fills up. Window size doesn't

<!-- SC_OFF --><div class="md"><p>The idea is simple, if one agent is good then multiple agents working together must be better. In practice coordination between agents is an