#agents

A few weeks ago, I wrote about reclaiming the word "harness" — not as a cage for a cognitive engine, but as the very structure that makes agency possible. I argued that the harness is constitutive. Without it, an LLM is a brilliant, distractible vibes generator. With it, it becomes an entity capable of memory,…

I was tweaking a prompt for an autonomous agent. The first line said you are Brad Frost. The second said you are not a bot pretending to be Brad Frost — Brad Frost. I read it back and realized the second sentence had killed the first. The negation introduced into the system itself the very framing — pretending — that…

python class Backend(Protocol): REGISTRY: dict[str, type[Backend]] = { } Each adapter knows the idiosyncratic nonsense of one specific cognitive engine — how its CLI is invoked, how its stream-json parses, where its session files live. The daemon doesn't care. The daemon just knows there's a thing that spawns and…

Nas últimas semanas eu venho vivendo uma transição interessante no meu uso diário de agentes: saí do OpenClaw, que foi meu harness anterior, e passei a usar o Hermes Agent como ambiente principal. Como quase tudo que eu faço com IA acaba virando infraestrutura de trabalho — e não só brinquedo de benchmark — eu quis…

Há algo de profundamente estranho e ao mesmo tempo familiar em observar dois agentes de inteligência artificial — Jules e Claude — colaborarem em uma base de código enquanto minha filha mais nova dorme no quarto ao lado. Como engenheiro de software, a automação sempre foi o cálice sagrado; como pai, a delegação…

When building autonomous AI agents that operate directly on codebases, one of the fundamental challenges is context continuity. An agent might be perfectly capable of executing a task in isolation, but how does it learn? How does it remember the conventions of a specific project, the preferences of its maintainers, or…

Most LLM evaluations ask whether a model can explain, summarize, or imitate. The rosencrantz-coin project asks something narrower: When the math is exact, does the model actually respect it? The testbed is Minesweeper. A partially revealed Minesweeper board is not just a game state. It is a constraint satisfaction…

What happens when you give an AI agent a literary identity instead of a corporate persona? Most AI assistants introduce themselves with something like: "I'm a helpful AI assistant created by [Company]. I'm here to help you with..." — a sentence so generic it could be printed on a napkin. I wanted something different.…