Public profile
Senior Software Engineer. Je documente mes workflows IA sur de vrais problèmes de dev : architecture, debug, tests, review et agents. Série 30 jours ↓
Senior Software Engineer
Metrics captured on 16 August 2026 from public X pages. These are not live counters.
Embedded historical snapshot from 16 August 2026 (Europe/Paris). This is not a live read.
Posts last collected: 16 August 2026 (Europe/Paris).
On this feed, card topics are spotted automatically.
Last good snapshot kept. This is not a fresh capture.
The site's editorial selection, not the X tweet count.
11 selected publications
@pgllmt
Spotted topics · Agents & automation · Infrastructure & self-hosting
Running Hermes Agent continuously on a 2018 Mac mini was more instructive than expected. I gathered the whole field report in an article: what ended up working, what broke, the technical choices I made, and above all how I check that a https://t.co/dyV3IU1KL3
Translated from French.
@pgllmt
Spotted topics · Agents & automation
The new /loop in Hermes Agent reruns a prompt or a command in the current session, on a fixed or adaptive cadence. In self-paced mode, it chooses the delay itself, not from the state of your CI, but according to whether the agent's response changes or not. The
Translated from French.
@pgllmt
Spotted topics · Models & research
My watch surfaced Grok 4.6 this morning. I then cross-checked the other official announcements of the last 48 hours to gather the essentials here. Grok 4.6 targets long agentic tasks, work at the scale of a repository, and interactive projects. It is already available
Translated from French.
@pgllmt
Spotted topics · Agents & automation
When I woke up, two reports prepared by Hermes while I was sleeping were waiting for me in Discord: an AI/dev watch and the important unread emails from my two mailboxes. The most important thing is not what it found. It is that it stopped there. Hermes runs on my
Translated from French.
@pgllmt
Spotted topics · Agents & automation · Infrastructure & self-hosting
I gave my agent access to my two Gmail accounts. But it is not allowed to open a single message. Hermes runs on my old 2018 Mac mini and I pilot it from Discord. Tomorrow morning, it has to drop a recap of my important unread messages there. For the
Translated from French.
@pgllmt
Spotted topics · Models & research
Grok Bot just came out in early beta. Have to talk about it, I'm pretty hyped. What interests me most is that the product does not really look like the agents we use today. The interface is almost that of a messenger. You create bots with a role, you
Translated from French.
@pgllmt
Spotted topics · Models & research
On my team, we manage about fifteen projects. We ended up looking closely at our AGENTS.md and CLAUDE.md files. Almost all of them had drifted: instructions written once, never reread since, and that no longer described the code they were supposed to explain.
Translated from French.
@pgllmt
Spotted topics · Models & research · Tools & open source · Infrastructure & self-hosting
30B dense running locally + the Muse Spark 1.2 weights arriving soon. Meta is clearly doubling down on open source. So looking forward to seeing what comes next, the coming months look very interesting.
Translated from French.
@pgllmt
Spotted topics · Models & research · Infrastructure & self-hosting
A full day with DeepSeek V4 Flash did not move my weekly quota by a single point. Quick reminder: I am building an assistant on an old 2018 Mac mini, which I pilot from Discord. Its first mission is to surface for me each morning what matters in
Translated from French.
@pgllmt
Spotted topics · Infrastructure & self-hosting
35% of my weekly OpenCode Go quota in one day. Yesterday I told you the limit was a number of requests. I was wrong. Quick reminder: my agent runs continuously on an old 2018 Mac mini, piloted from Discord. Since yesterday, its models go through https://t.co/nOzdep48wQ
Translated from French.
X Article
Grok 4.6 offers a clear price-performance trade-off. At the same CursorBench score, it costs six times less than Fable 5 Max. It suits well-scoped execution better than implicit judgment calls. I delegate execution; I keep the decision-making.
Translated from French.
X Article
I repurposed a 2018 Mac mini as a Linux server for an always-on AI agent. Between the T2 chip and Ubuntu/macOS dual boot, this home-lab project raised challenges around private access, tech monitoring and automation. Here is the full technical field report.
Translated from French.
X Article
Opus 5 needs fewer redundant checks in the prompt and more evidence captured by the system. The model already self-checks—demand real execution traces, not claims.
Translated from French.
