An `AGENTS(.)md` (or equviliant) is the highest configuration point for agents. It's injected into every conversation. But research shows that doing it wrong actively hurts performance. Here's how to do it right, backed by data.
Less Is More:
- Auto-generated files reduce success rates by ~3% while increasing inference cost by over 20%
- Human-written files only marginally improve performance on benchmarks (~4%)
- Stronger models don't generate better context files.
- Codebase overviews in files don't help agents navigate faster.
- LLM-generated files are redundant with existing docs.
- Instructions ARE followed. Agents respect file instructions, but unnecessary requirements make tasks harder.
What to Include:
- WHAT: Your tech stack, project structure, and what each part does. Critical for monorepos.
- WHY: The purpose of the project and its key components. Help the agent understand intent, not just structure.
- HOW: How to build, test, and verify changes. Include non-obvious tooling (e.g., `uv` instead of `pip`, `bun` instead of `npm`). Tools mentioned in AGENTS(.)md get used 160x more often than unmentioned ones.
What NOT to Include:
- Detailed codebase overviews or directory listings. The paper found these don't help agents navigate faster, and agents can discover structure themselves.
- Code style guidelines. Use linters and formatters instead, they're faster, cheaper, and deterministic.
- Task-specific instructions that only apply sometimes. - Auto-generated content. Don't let the agent write its own without review. The data shows this hurts more than it helps.
How to Structure It:
- Keep it short. General consensus is <300 lines; HumanLayer keeps theirs under 60 lines. Every line goes into every session, make each one count.
- Use progressive disclosure. Don't put everything in Instead, keep task-specific docs in separate files (e.g., `agent_docs/running_tests.md`) and list them with brief descriptions so the agent reads them only when relevant.
- Pointers over copies. Reference `file:line` locations rather than embedding code snippets that will go stale.
- Write it yourself, deliberately. A bad line cascades into bad plans, bad code, and bad results across every session.
Sources below
I built Security Agent for #GoogleWorkspace.
It is using Agent Development Kit (ADK), Gemini 3 Flash and Policy API (OpenAPI schema)
Check my notes and full-code
https://t.co/D8F7q4YT11
#AISprintH2@workspacedevs @expert
Wow! ๐ Honored to win the "AI Personality of the Year" 2025 in the ๐ Education category at the Czech Association of Artificial Intelligence event.
This award is voted by the publicโthe people I've trained. That makes it so special. ๐ Thank you!
Finally had some time to work on a Gemini CLI feature I've been meaning to get to for a while now.
Full control of MCP servers from the command line. Devs shouldn't need to learn about settings.json ๐
โ gemini mcp add <name> <commandOrUrl> [options]
โ gemini mcp remove <name>
โ gemini mcp list
I'll link the draft PR below in the comments if anyone wants to provide some initial feedback ๐๏ธ
@ntaylormullen Check out this ๐ฅ๐ฅ๐ฅ
Google DeepMind Team Close to Solving One of the Seven Millennium Prize Problems with AI.
If the team nails the proof, fluid mechanics gains a solid foundation and AI earns bragging rights as a serious research partner.
For these 7 unsolved math puzzles, there's a $1Mn prize for each correct solution, so researchers nicknamed them the Millennium Prize Problems.
The Navier-Stokes formulas, track how fluids move, still hide a scary question: can smooth flow suddenly erupt into an infinite spike, like a surprise tsunami inside calm water?
Weather forecasts, plane design, and even blood-flow analysis all trust that spike never appears, so proving it one way or the other truly matters.
Google DeepMind has lent 20 engineers plus its heavy compute to Javier Gรณmez Serranoโs small math crew. They train neural networks on past proof tricks and fluid simulations, letting the code spot patterns that hint where a flow might explode.
Those hints guide the humans to the exact region and time where a singularity could form inside the Navier-Stokes equations, something pencil work alone kept missing.
AI also polished Thomas Houโs older Euler-based test. The team fed that simulation back into a machine-learning loop, tightened the grid, and the network marked the first flicker of instability.
With that refined map, analysts can focus their rigorous estimates on one narrow zone instead of swimming through the whole ocean of possibilities.
---
english .elpais .com/science-tech/2025-06-24/spanish-mathematician-javier-gomez-serrano-and-google-deepmind-team-up-to-solve-the-navier-stokes-million-dollar-problem.html
My key takeaways from #GoogleIOConnect
- use NotebookLM to prepare prompts (with grounding to docs) for Firebase Studio
- MCP toolbox connect to your database
- GenAI Processors,lightweight Python library for parallel content processing
#BuildWithGemini
-
It's 2025, and Iโve finally updated my Python setup guide to use uv + venv instead of conda + pip!
Here's my go-to recommendation for uv + venv in Python projects for faster installs, better dependency management: https://t.co/slD3Vxpbpd
(Any additional suggestions?)
๐ค Can AI Watch YouTube video and Take Notes for You? Automating Insights with #Gemini 1.5 Pro & #GoogleAppsScript
Non-stop #YouTube research bot, which find new related videos, watch it and taking notes into the same Google Docs
https://t.co/16IBmFKqvF
#AISprint#GoogleCloud
๐ If you haven't already tried it: @googlechrome now has an embedded version of Gemini - and specialized guides for developers detailing how to use other local models within your browser.
football AI code is finally open-source
- player detection and tracking
- team clustering
- camera calibration
I still need to work on README; don't judge me on that
code: https://t.co/VK0RQFX22z
As part of @Google's 1,000 Languages Initiative, a commitment to support the 1,000 most spoken languages, & w/help of our PaLM 2 LLM, we're adding support for 110 new languages (spoken by 614M people) to Google Translate (now supporting 243 languages). ๐
https://t.co/DRiw03V0wu
As the Official Data Analytics Partner of the England Team, find out how @FA is using Google Cloud AI to find the next generation of England football stars. #GoogleCloudAI
Read the blog: https://t.co/9efsrtWB3D