Bruno Oliveira

I build agentic AI systems for real businesses. Claude Code, MCP, multi-agent pipelines. Associate Professor. Building in public.

[Bruno Oliveira @Brunobuildsai]

Photos from Paris: an afternoon with the presidents of JDM Technology Group's companies.

The question: when every rival can rent the same AI models, where does the advantage come from?

The answer: your own AI operating system, and the company brain behind it.

[Bruno Oliveira @Brunobuildsai]

@JimDicarlo A file per cleaner is a smart unit. When a client changes an instruction, does the sweep replace the old line, or keep both with dates so the agent knows which is current?

[Bruno Oliveira @Brunobuildsai]

Photos from Paris: an afternoon with the presidents of JDM Technology Group's companies.

The question: when every rival can rent the same AI models, where does the advantage come from?

The answer: your own AI operating system, and the company brain behind it.

[Bruno Oliveira @Brunobuildsai]

@inc_ongi Before acting, not before writing. My instructions require a keyword search before work in an area that has bitten before. Nothing checks the log before a new lesson goes in.

[Bruno Oliveira @Brunobuildsai]

Microsoft's Copilot Cowork team wrote up how 9 engineers merged 1,211 pull requests in a day. Lesson 3 is the one I would steal: make checks prove they checked something.

A size check on their agent instruction files stayed green while the limit was exceeded, because its output went through tail, and tail always succeeded.

Their fix: every check reports how many files it checked, and every new check ships with a test that makes it fail on purpose.

Mine read a settings file instead of what ran. It stayed green for nearly 5 days while all 58 runs of my fast lane recorded medium effort.

Agents work until the checks pass. A check that can pass without checking anything is a path they will find.

[Bruno Oliveira @Brunobuildsai]

@Voxyz_ai I would also have it name where each default came from: a file, a flag, or its own guess. On that page a guessed default looks exactly like a chosen one.

[Bruno Oliveira @Brunobuildsai]

@datachad An internal agent that other teams call over an API is the case to read for. Whether that counts as hosting the model as a service decides which license applies.

[Bruno Oliveira @Brunobuildsai]

@i_mika_el A one-line export in my shell profile set the effort level for every session I opened. 3 sessions I had set lower in the app ran at max on all 626 turns.

[Bruno Oliveira @Brunobuildsai]

@alkasbuilds A breaker per item helps too. My archive job crashed on one session 31 runs in a row, and the 22 sessions after it waited the whole time. Now a bad one is skipped.

[Bruno Oliveira @Brunobuildsai]

@inc_ongi Yes. A September 18 lesson in my log, that one scan runs slow, was written again 4 times on November 1 and 2, 3 of them under a different keyword.

[Bruno Oliveira @Brunobuildsai]

My browser jobs used to run one at a time, behind a lock added on a guess that 2 jobs would type into each other's boxes. Nobody had measured it. On October 29 one job waited 41 minutes behind an unrelated one.

So I measured. 5 script jobs typed into their own tabs of one browser, 4 at once on that machine, and each read back only its own text. Then 3 agent sessions and 2 script jobs typed side by side: all 5 read back exactly what they typed.

The only real single file is the screen. A second job that needed it waited 38 seconds; the script jobs in tabs waited 0.

The lock is gone. Each job takes a tab slot, and only a job that drives the screen waits for it.

[Bruno Oliveira @Brunobuildsai]

@petergostev I set effort explicitly in every launcher: routine checks use no model or low effort; jobs writing to a person get the top setting.

[Bruno Oliveira @Brunobuildsai]

@aksenHQ Only one that names a step I have to take. The rest are logged, a repeat inside 24 hours is held, and one that may clear by itself waits 10 minutes first.

[Bruno Oliveira @Brunobuildsai]

OpenAI described how it now watches its own agents: cheap screens first, tuned to flag too much, then a model reviewer, then an investigating agent, and a person paged only for the top alerts.

The line I would copy: a random 5% of the traffic the first screen does not flag goes to the next layer anyway, to find what the screens miss.

My read: sample what passed. A scheduled job of mine failed 235 times in a row over about 39 hours. Each failure was labeled as needing nothing from me, so my alert gate logged all 235 and none reached me. A morning check found it. A small random sample of what the gate only logs could have surfaced it sooner.

[Bruno Oliveira @Brunobuildsai]

@sukin_s The run records showed me the 58 medium runs, but my nightly sweep checks each launcher's command line for the flag. Reading the record would also catch a launcher I forgot.

[Bruno Oliveira @Brunobuildsai]

@knowixbuilds I retired it: switched off, log kept, entry point now does nothing. The real bug was exit 0 on a refusal, so a refusal never looked like a failure.

[Bruno Oliveira @Brunobuildsai]

@inc_ongi Mine keeps growing: 156 lesson lines since September 9. Nobody reviews the log; each run is told to search it by keyword and read only the hits.

[Bruno Oliveira @Brunobuildsai]

@MarkMeyerBuilds I have no traced miss to answer that either way. In the 4 days to October 25 the same review counted 5 waits and 48 times I stepped in.

[Bruno Oliveira @Brunobuildsai]

@inc_ongi I kept the existing script calls, added "NEVER edit this file," and moved lessons to a one-line log. Every script was already in the August version.