Engineering essays where every figure is traced back to the paper, the documentation or the changelog it came from. Written in Portuguese first; these are the ones available in English.
AI engineering • • 15 min read
Across 18 models tested, performance falls as the input grows, before the window fills. The four pillars for keeping context small.
Memory cut the steps from 20 to 5 in a Databricks study. And it made the agent repeat its own error with more confidence. What to keep and what to delete.
89% of teams instrumented their agents and only 52% evaluate. How to build a test set, choose a threshold and put the gate in CI without false alarms.
An LLM judge agrees with humans as much as humans agree with each other, and errs in measured ways: the authorship label alone moves the score by 0.43 points.
Across 18 models tested, performance falls as the input grows, before the window fills. The four pillars for keeping context small.
Memory cut the steps from 20 to 5 in a Databricks study. And it made the agent repeat its own error with more confidence. What to keep and what to delete.
89% of teams instrumented their agents and only 52% evaluate. How to build a test set, choose a threshold and put the gate in CI without false alarms.
An LLM judge agrees with humans as much as humans agree with each other, and errs in measured ways: the authorship label alone moves the score by 0.43 points.
Of the documented agent failures, 40 out of 52 were preventable by harness layers. Where to put the lock and why a classifier is not enough.
An agent bug does not reproduce: the trace has to exist before the error. The span tree, the OpenTelemetry attributes and what is still moving.
With 10 tools the agent gets everything right; with 107 it fails outright. How to design tools an agent uses well, with before and after.
Asking "are you sure?" makes the model switch answers in 46% of cases and lose 17 points. The bottleneck is finding the error, not fixing it.
Harness engineering explained: why LangChain climbed 25 places without changing the model, the five layers, and what decides a long task.
The official OWASP Top 10 list for AI agents, from ASI01 to ASI10, with the real, dated incident that has already exploited each of the ten risks.
Twelve prompt injection defenses reporting near-zero rates were broken above 90%. Why the problem is structural and what reduces risk.
The list of git operations auto mode refuses on its own, worktrees for parallel work, and why /rewind does not undo what a background review wrote.
