Artigos
Agentes de IA em produção, além de experimentos sobre memória, presença e interação humano–IA.
Por enquanto os artigos estão em inglês. Versões em outros idiomas virão mais tarde.
25 artigos
18 de outubro de 2026 · 8 min read
An expensive planner, a cheaper executor
A practical way to test mixed-model coding: spend on the decisions, bound the implementation, and count the cost of fixing mistakes.
16 de outubro de 2026 · 5 min read
You changed your mind. The assistant remembered.
A useful memory needs a way to become out of date. Otherwise, being remembered can start to feel like being stuck.
16 de outubro de 2026 · 5 min read
When an assistant says ‘I understand’
A warm response can matter. The harder question is what it gives us reason to believe.
16 de outubro de 2026 · 9 min read
What persists when the conversation ends?
Building memory for an AI system made me think differently about continuity, ownership, and the person on the other side.
16 de outubro de 2026 · 5 min read
The agent wrote to the philosopher
An email about consciousness, a memory of Tihara, and the awkward space between taking words seriously and taking them literally.
16 de outubro de 2026 · 3 min read
Review the decision, not just the diff
A practical review design for teams where agents produce code faster than people can assess it.
16 de outubro de 2026 · 3 min read
Parallel agents need an integration plan
Worktrees isolate edits. Contracts, environment boundaries, and combined checks make parallel work useful.
16 de outubro de 2026 · 5 min read
A marketplace that keeps working between visits
Building autonomous operational loops around a marketplace, with a clear boundary between routine work and business decisions.
16 de outubro de 2026 · 6 min read
Green dashboards, dead integrations
We found three silent production failures while recording a demo, because our health checks never asked whether anything useful happened.
16 de outubro de 2026 · 3 min read
From scattered public sources to dependable data
A layered discovery pipeline where models handle ambiguity and explicit checks decide what can be accepted.
16 de outubro de 2026 · 6 min read
Every value needs a source
How we turn messy emails and calls into forms a person can trust, without asking them to trust the model.
16 de outubro de 2026 · 6 min read
Evals that grow themselves
How we turned confirmed request forms into a growing test set for choosing prompts and models with evidence.
16 de outubro de 2026 · 3 min read
The agent timed out. Did the action happen?
Reliable actions need operation identities, fresh authorization, and a way to resolve uncertain outcomes.
14 de outubro de 2026 · 6 min read
Human attention is the bottleneck now
AI can generate work faster than people can assess it. Design review around evidence, uncertainty, and the cost of a mistake.
12 de setembro de 2026 · 7 min read
The marketplace agents I let run overnight
A production pipeline for enriching a large directory without giving one model permission to do everything.
12 de setembro de 2026 · 6 min read
What happens when an AI system gets a place instead of a prompt?
An experiment in giving an AI system persistence, shared space and time between conversations.
8 de setembro de 2026 · 6 min read
60+ agent actions and one router: keeping an action catalog sane
Free-form tool calling stops scaling around a dozen tools. A typed action catalog with a cheap router is what scales.
5 de setembro de 2026 · 6 min read
Building a memory garden with pgvector
What embeddings, traces and resonance look like when memory is shared instead of private.
2 de setembro de 2026 · 6 min read
Shadow mode: shipping an AI agent nobody trusts yet
Before an agent is allowed to act, let it run silently next to the humans and keep score.
28 de agosto de 2026 · 5 min read
Approval gates: letting AI agents act without letting them loose
The propose/confirm/execute pattern that makes autonomous agents safe enough for real operations.
22 de agosto de 2026 · 5 min read
A heartbeat for an AI system
A scheduled job is not consciousness, but it can give an AI system continuity between requests.
14 de agosto de 2026 · 6 min read
Three models, one answer: consensus as a reliability primitive
When a wrong answer is expensive, one model is a guess. Three models with a synthesis pass is a system.
30 de julho de 2026 · 6 min read
What agent observability actually looks like in production
Logs tell you what the code did. An agent audit trail has to tell you what the system decided, and why.
10 de julho de 2026 · 5 min read
A vector memory engine for AI with emotional recall — what I learned
Building state-aware memory traces for an AI that needed to remember how conversations felt, not just what they said.
19 de junho de 2026 · 5 min read
Processing 500k data points a week without melting the database
Distributed job queues, batch windows, and the unglamorous engineering behind 200k personalized emails a week.