Antoine Bénéteau

Founder @ Goodvest | Impact investing | AI Engineer

[Antoine Bénéteau @_abeneteau]

What can a test dependency read inside your agent's sandbox?

A Git token in the clone URL ends up in .git/config. Cloudflare's visual guide shows the fix: add credentials outside the sandbox, only to approved HTTPS requests.

https://t.co/kpobrRnQ0h #AI #Security

[Antoine Bénéteau @_abeneteau]

A tool call can fail while your CI stays green.
This MCP Inspector guide shows how to catch that: call a read-only tool, assert its output, and preserve the failure code through shell pipes.

Script + GitHub Actions example:
https://t.co/Lbd76BpSwM
#MCP #AI

[Antoine Bénéteau @_abeneteau]

What actually turns an LLM into an agent?

Build one in ~60 lines of Python and watch it choose commands, read the results, and decide what to do next.

This hands-on guide makes the whole loop visible. Use a sandbox.

https://t.co/Mxvs4om1R1

#AIAgents

[Antoine Bénéteau @_abeneteau]

An agent's own notes can contain fake user approval.
Malicious skill chains triggered unwanted actions in 512 of 690 controlled SkillsBench runs this way.
Check authorization when the tool acts. A generated note saying “approved” should never be enough. #AI #Agents

[Antoine Bénéteau @_abeneteau]

An agent's own notes can contain fake user approval.

Malicious skill chains triggered unwanted actions in 512 of 690 controlled SkillsBench runs this way.

Check authorization when the tool acts. A generated note saying “approved” should never be enough. #AI #Agents

[Antoine Bénéteau @_abeneteau]

Adding a manager agent made a team worse.

In one simulated merge queue, Opus 5 reached 65% of optimal release value. With a manager: 38%. With a manager, explicit priorities and a PR to drop: 95%.

An agent org chart still needs a way to say no.

#AI #Agents

[Antoine Bénéteau @_abeneteau]

@KomatsuSo68 Exactly. Run both on the same tasks and count retries and human intervention too. A cheaper first attempt means little if someone has to rescue the workflow.

[Antoine Bénéteau @_abeneteau]

An AI agent wrote the code that chooses which AI model it uses.

In Pi 1.0’s demo, Opus plans, GPT implements, and a router switches between them.

I'd like to see that tested against one model doing the whole job. Does the extra complexity pay off? #AI #Agents

[Antoine Bénéteau @_abeneteau]

Hugging Face's Serge works the bug backlog overnight: 29 fixes merged into Transformers in 80 days, with GPU verification and human review. Recent inference cost: ~$43 per merged PR. Which maintenance job would you hand to an agent first?

#AI #Agents

[Antoine Bénéteau @_abeneteau]

Ai2 reports scaling Olmo-core 3 from 4.6B to 47B parameters with under 5% lower training throughput. Only ~3.2B activate per token. The catch: larger tests used random routing. Promising systems work; model quality is a separate question.
#AI #OpenSource

[Antoine Bénéteau @_abeneteau]

H Company reports 85.2% on the original OSWorld benchmark at $0.08 in API costs per task with Holo4 27B. It also published 7,366 agent runs with actions and screenshots. You can inspect the failures before trusting the score.

#AI #Agents

[Antoine Bénéteau @_abeneteau]

Google says Gemini 4 Argon agents analyzed profiling data and found ways to free over 300 TiB of memory across its data centers.

Could performance engineering become cheap enough to run continuously?

#AI #Engineering

[Antoine Bénéteau @_abeneteau]

“We’ll add it to the roadmap” gets awkward when a customer’s AI can build it before your next sprint.
If software can adapt on demand, which SaaS features are still worth paying for?
#AI #SaaS

[Antoine Bénéteau @_abeneteau]

@Sametheus @AnthropicAI Fair point on the harness. But the eval did catch the tailures. The open question is whether extra reasoning caused the drift or exposed a weakness in the setup.
These two cases don't settle that

[Antoine Bénéteau @_abeneteau]

@mehdizare Exactly. What’s interesting here is that the extra compute didn’t just cost more, it produced a worse outcome on this benchmark

[Antoine Bénéteau @_abeneteau]

Even AI can overthink.
Anthropic reports Sonnet 5.5 scored worse at Max reasoning than Xhigh on FrontierCode. In two reviewed cases, extra review agents caused a timeout or edits outside the task.
#Claude #AIAgents
https://t.co/HAohCwU0ym

[Antoine Bénéteau @_abeneteau]

Cloudflare got a Doom demo running in a terminal via its browser for AI agents.

Kitesurf runs page code on Workers and displays it in your terminal, so you can inspect what your agent sees.

Free beta, with account limits. #AI #WebDev
https://t.co/ENu1ZqV5th

[Ankit @BytesWithAnkit]

I literally can't focus on anything but Dario here 😭

[Kylie Robison @kyliebytes]

i have watched this video 20 times