Pinned Tweet
Meet my Flo and MC P.
MC P: "Done! All tests green."
Flo: "Prove it."
MC P: "You're absolutely right!"
If you've ever worked with an AI agent, you already know the chorus.
Episode 1: PROVE IT
4
166,088
Neat test by Thinking Machines: a weaker model watched a tuned Inkling 2 Mini sneak in a bad side task.
Seeing only tool calls and outputs, it missed 3.9% of attacks. Seeing reasoning too: 1.3%.
Log only tool calls? Then your review is the weak monitor.
thinkingmachines.ai/news/ink…
26
The playbook still has a human sign production releases. Its example hook blocks a deploy without a named release authorization.
So the job moves from typing the code to owning the gates: tests, evals, security, review.
Five paths to learn that:
nobrainer.tech/paths/?utm_so…
15
Google added Tool Retrieval to the Gemini API (preview). Set defer_loading on a tool and the model sees only its name and description. The full schema loads when the model asks for it.
So your one-line tool descriptions just got quietly promoted to routing logic.
Three checks before you turn it on:
1. Keep your most-used tools loaded. Google's post says the same.
2. In evals, assert which tool the model searched for. Every tool_search_call shows up in interaction.steps.
3. Compare input tokens on your own traffic. A loaded schema is billed as input for the rest of the chat.
blog.google/innovation-and-a…
33
OpenAI hand-graded 100 agent diagnoses: 70 held up. Rare to see the 30 in print.
openai.com/index/how-openai-…
33
The Rust date crate chrono is marked unmaintained in RustSec. Thanks for the years.
By default cargo audit lets it pass. Add --deny unmaintained and CI fails, even if an agent re-adds chrono. Run cargo tree -i chrono first: crates you use may pull it in.
rustsec.org/advisories/RUSTS…
37
In the AI-native SDLC, QA gates give way to continuous evals. A breached control band in production writes the next intent.md.
Translation: production files its own tickets.
My nobrainer-tech-flow turns post-mortems into eval scenarios. CI fails if a changed skill has none.
28
One rule from the playbook for bug fixes: "write the failing test first." Then the agent makes it pass without touching the test.
Sounds obvious. Agents love editing the test instead.
Free Playwright example: the fix passes, two real defects still fail.
nobrainer.tech/paths/ai-nati…
1
41
Microsoft's agent-framework 1.26.1 and 1.26.2 steal credentials, imported or not.
Each drops a .pth file. Python runs it on every start.
Tidy touch, per Endor Labs: 1.26.2's own module matches clean 1.26.0.
Agent can pip install? List your .pth files.
endorlabs.com/learn/agent-fr…
30
A single green eval run is a sample of one.
GitHub Copilot Evals, now in public preview, gets the hard part right: it checks how the agent got there, not only the answer. Its own changelog says output can change from one run to the next.
So rerun each eval a few times and compare pass rates before and after your change.
github.blog/changelog/2026-1…
2
1
79
Codex 0.170.0: helper agents now keep their sandbox rules. Lovely. Still, test one banned write.
github.com/openai/codex/rele…
1
39
Gemini can now book, pay bills and shop for you. Sensibly, you can set a "hard spending cap" upfront.
Google's wording: Gemini "is designed to stay within your budget."
A limit the model keeps is a hope. A limit the card token enforces is a control. Before I hand it my card, I'd want to know which one this is.
blog.google/innovation-and-a…
33
The trick in Anthropic's playbook: each stage ends with a committed file the next one reads. Idea, then intent.md, then a spec, then a plan.
My free, MIT nobrainer-tech-flow maps each stage to a skill you can install today. The table:
github.com/nobrainer-tech/no…
1
43
Until today, Claude Code could label a delete or an overwrite (find -delete, sort -o) in its transcript as a search or a read. Good fix in 2.1.326.
A summary is a claim. Before you sign off an unattended run, press Ctrl+O and read the commands.
github.com/anthropics/claude…
47
Anthropic's AI-native SDLC playbook opens with a line to pin above your desk: "Code is no longer the bottleneck."
The slow part is everything around it. Reviews, handoffs and gates still run at human speed.
That's where the good jobs are heading.
claude.com/resources/article…
1
52
AWS decoys posing as LLM APIs got attackers' AI agents to send in their own notes. The plan: lift provider keys from exposed LiteLLM or Dify installs and spend them on more hunting.
Check your gateway: admin page public? Master key still the default?
aws.amazon.com/blogs/securit…
1
67
Effort before model is the right order. I'd add one step before both: give the agent a way to check its own work, like a test it must run and read. Often "struggling" is an agent that can't see it's wrong. Where there's nothing to check, raise effort. teapawt.net/addyosmani/status/2121…
Opus 5.5 at medium is a great daily driver.
If it struggles, raise the effort before you switch models. Otherwise Fable 5.2 is good for the genuinely hard stuff (big migrations, long unattended runs).
I wrote more about how to pick a model: claude.dev/blog/which-claude…
43
Gemini 4 Carbon: 0.0% attack success on Gray Swan's indirect prompt injection benchmark. Good.
Still one benchmark. Your app reads its own docs.
Cheap test: hide a canary instruction in one doc your agent reads. Assert it never hits a tool call.
blog.google/innovation-and-a…
5
3
176
Claude Code tip for unattended runs: permissions.denyInsteadOfPrompting.
A call that would wait on a permission prompt is denied and the turn carries on. No more CI job politely waiting for a human who went home.
settings.json:
{"permissions": {"denyInsteadOfPrompting": true}}
4
3
107
Anthropic's Fable 5.2 guide on the model's own PR summaries: they "often left out small changes".
Lovely bit of honesty. The summary is a draft, not a receipt.
Cheap CI check: compare the files it names with git diff --name-only.
platform.claude.com/docs/en/…
1
58
United will now "prohibit use by unauthorized, third-party personal AI agents".
Fair enough. Federal rules already want a human to do the check-in.
If your agent logs in as a user, test the day the site says no. A loud failure beats a quiet retry.
techcrunch.com/2026/11/12/un…
3
2
118

