Bruno Oliveira (@Brunobuildsai) header
Bruno Oliveira (@Brunobuildsai) profile picture

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

Bath, United Kingdom
Joined July 2024
Photo by Bruno Oliveira (@Brunobuildsai)Photo by Bruno Oliveira (@Brunobuildsai)Photo by Bruno Oliveira (@Brunobuildsai)Photo by Bruno Oliveira (@Brunobuildsai)Photo by Bruno Oliveira (@Brunobuildsai)Photo by Bruno Oliveira (@Brunobuildsai)Photo by Bruno Oliveira (@Brunobuildsai)Photo by Bruno Oliveira (@Brunobuildsai)Photo by Bruno Oliveira (@Brunobuildsai)Photo by 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.
4
5
175
On November 10, 3 workers of one batch image job on my main Mac grew to 207, 171 and 167 GB: 545 GB together, on a machine with 512 GB of memory. It ran out. My guard back then could only pause jobs, and a paused process keeps every byte it holds. The fix that afternoon: a fresh worker for every block of frames, its memory read 4 times a second, stopped at 32 GB. The 65 blocks that job then ran peaked at 11.3 to 25.8 GB each. None reached the cap. On November 12 a second guard ended 4 other jobs, each holding 113.7 to 117.5 GB.
4
2
72
Anthropic's new guide to always-on agents starts with an admission: the agent spends most of its time asleep and wakes periodically to do work. Around that loop it names 5 parts: an event inbox, timers, memory, permissions and an outbox. For credentials it recommends a proxy outside the agent's environment, so the agent's requests go out without them. The line I would underline is in the outbox section: start quiet, then earn more of the user's attention over time. My own alerts go through one gate. In its first 3 and a half days, 24 alerts reached it and 7 reached me.
4
3
207
Who may edit an agent's instructions? In my system the answer is split in 2. The system applies a change by itself in only 2 cases: it removes a restriction one of its own runs wrote, or it automates a step I was doing by hand. Everything else waits for my decision. A review of the moments I had to step in ran 9 times between October 15 and November 15 and proposed 66 changes. 35 are applied, most of them under that standing rule. Each applied change has its own folder: the files as they were before, whose approval it rests on, how it was checked and the exact undo. No undo has been used so far. Where would you draw that line?
2
1
108
Sam Nickless runs Gilbert + Tobin, an Australian law firm, and two of his AI habits are worth copying. He gave a custom GPT his writing, feedback he had received, firm context and his working style. His direct reports test their thinking against it before coming to him. It does not decide or speak for him. He also checked the bill. Updating files from a spreadsheet, making PDFs and bundling the annual review for 120 partners took minutes. The usage dashboard showed about $62. His feedback now sits where the team can reach it, and he knows what a run cost.
3
1
1
247
A folder of about 29,000 small records on my Mac kept being read by the search indexer, though nothing in it had changed. A 6-second trace on November 4 counted 34,294 file calls by the indexer on that one folder. My own sync between 2 Macs runs every 2 minutes, and my notes from that afternoon say each pass stamped every copy as changed again. The fix that afternoon: the sync and my scheduled backups now leave unchanged files alone, and the folder is out of indexing. Read again on November 9: none of the 29,082 older records had been stamped since the fix. The 410 new ones were each stamped once, on arrival.
1
25
I would borrow Curative's second reader before giving an agent more authority. In Anthropic's write-up, their contracting agent, Gwen, writes out why it wants to go above its rate template. A Sonnet reviewer clears about two-thirds of those requests. The rest go to the CEO, who says they would have reached him anyway. They added a Haiku judge after Gwen copied providers on internal notes about how high it could go. It reads every outgoing message, and a blocked one comes back with the reason so Gwen can rewrite it. Every few days the CEO also has Claude Code sample the last 10,000 runs, flag anything weird and open a pull request. That habit I would copy too.
3
2
77
Overnight on November 9 I built my own search over my whole file store. No model in the search path. SQLite FTS5, a porter stemmer, prefix indexes. By 11:50 that morning it listed 137,981 files and read inside 49,071 of them. During the build, a search for the single letter "a" took 15.9 seconds, because the built-in snippet function re-stemmed whole documents. I now cut snippets from the stored text myself and skip a lone letter while it is being typed. After that fix, 12 overnight test searches came back in 13.7 to 36.6 ms on the server. What is the slowest query your own index ever served?
1
41
Franklin Templeton's first try at AI was the obvious one: put the data in a database and connect the models to it. By OpenAI's write-up, it did not produce the answers the teams needed. The models did not understand how the data fit together. Results changed when they added agents and a semantic layer: a structured model of what the firm's data means and how it is connected. The detail I would steal is the other one. Many tools learn what a user prefers and show more of the same. They want the opposite: people accessing a diversity of research and opinion. My read: describe the data before you connect it, then decide which user preferences the system should ignore. Which is harder to ship?
1
3
77
My agent jobs that drive a browser had left 78 tabs open on the laptop they work on. On November 2 I added a cleanup with one rule: when a job ends, close only the tabs that job opened, and keep any tab with typed text, a running transfer or unsaved work. Its pass over the old leftovers closed 22 and kept 15 it judged unfinished: 14 had text sitting in a form, 1 had a transfer running. I tightened the check, a final pass closed 3 more, and the laptop ended the evening at 41 tabs. The part I would copy: each job records the tabs it opens. The cleaner never guesses from a before-and-after count, so jobs running side by side do not close each other's work.
4
3
83
Anthropic's new post on agent memory has one small story I find more useful than its benchmark numbers. In one run, Claude trusted a note that put a landmark on the wrong side of a mountain and dropped off the same one-way ledge 3 times. Run again after a "dreaming" pass had checked the note against the session transcript and rewritten it, Claude got through. The numbers back it at scale: on a merger document set, 87% with a maintained wiki against 73% with search alone, on questions that join several documents. What I take for my own agents: notes they rely on need a scheduled check against the record.
7
5
220
For about 35 hours, the list on my phone that shows what is waiting on me stopped changing, and nothing said so. A script rebuilds it every 5 minutes. From 12:29 on October 24 to 23:10 on October 25, every rebuild froze on one line: a text-cleaning regex with a repeat inside a repeat, \n(\s*\n)+, which slows exponentially on a long draft. Each pass hit its time limit before it published, so the last good list just sat there. The fix was a plain pattern that does the same job in no measurable time. The bigger fix was a watcher with no model in it that ends a stuck rebuild, restarts it, and tells me only if that fails. A frozen feed looks exactly like a quiet day.
3
2
95
OpenAI's cyber team tested what their scaffolding was still worth. For an older model they packaged 103 exploitation techniques and prompted the agent to use them: success rose from 76.2% to 88.9% across 63 candidates. The newer model found and applied techniques on its own. Its light workflow beat the old scaffolded one, and adding the 103 back gave 1.5 points. They removed the layer: not worth the upkeep. Mine grew the other way. An hourly monitoring agent allowed to write lessons into its own instructions took them from 86 KB to 1.18 MB. None of it was new capability. I rebuilt them at 12.7 KB. When the model improves, test the scaffold again.
26
On November 2 my Mac stopped starting its scheduled jobs after a restart was begun and then canceled. A job that runs every 10 minutes went 79 minutes without a run. Starting a job on demand still worked, so I wrote a stand-in with no model: every 10 to 15 seconds it starts each job that is due. It made 65,733 starts in about 2 days. Its only ways out were a restart of the Mac or a stop file. The scheduler came back without a restart. The stand-in made its last start on November 4 at 10:43, then kept looping for 18 more hours until a nightly check stopped it. A stand-in should exit when the thing it replaces works again.
1
2
141
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.
1
58
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.
2
1
48
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.
2
2
66
In early October a watchdog on my Mac, a script with no model, told me every time one condition changed. The condition comes and goes by itself. From October 2 to 6 it changed 47 times and sent me 47 lines. None asked me to do anything. On October 6 I rewrote the rule: say it once per restart, the first time it holds, and log every change. Since then the condition has changed 343 times. The watchdog has sent 1 line about it. Twice more it reached the gate in front of my alerts, and the gate held both. A state that flaps belongs in a log. Which of your alerts ask you to do anything?
30
IDC, in technology research for more than 60 years, put its data inside Claude through an MCP connector. A CIO asks about procurement software and gets the top 3 vendors as an Excel scorecard, cited to IDC's research. The line I would copy: Claude gets the same data and definitions an IDC analyst works from, so a market share matches what IDC would publish. A vendor announcement, with no usage figures yet. My read: test the definitions before the connector. Take one published market share and check period, geography and denominator. A citation can point to the right document while the answer uses the wrong slice. What would you check first?
1
1
97
On October 3 a push connection on my Mac was reset 15 times in under 4 hours, and it never lasted past 32 minutes. I measured it: idle connections survived 4 minutes of silence and died by 6. I set the operating system to send a keep-alive after 120 seconds. The next connection held for 30 hours 53 minutes, and that was the story I was about to tell. An hourly check with no model kept reading the system log. In the 4 weeks since, it has logged 126 more resets, 61 of them in the last 7 days. The median reset now comes after about 3 hours. Better, and still unsolved. Without the check I would be quoting the 30 hours.
1
67