Ilan Chemla

Building AI agents at @nimble_search. Sharing what I learn along the way...
Also a Scuderia Ferrari fan on weekends.

[Ilan Chemla @ilChemla]

@svpino the more coding agents you run in parallel, the more code you auto-approve without reviewing.

You end up shipping garbage just to keep the agents printing.

Speed without review isn't speed. It's debt.

[Ilan Chemla @ilChemla]

💥https://t.co/dfcOv0xgzU

[Michael Grinich @grinich]

Today WorkOS is launching auth.md

An open protocol for agents to register for services on the web.

We're partnering with @Cloudflare and @Firecrawl as some of the first providers.

Why did we build this? And why now? 🧵

[Ilan Chemla @ilChemla]

@trq212 Cool! The auto recap when you re-attach is my favorite detail!

One line and I'm back in. No endless scrolling of the session.

Super useful 💪🏻

[Nimble @nimble_search]

The Nimble Plugin for Claude is going viral! 820k installs in the last 10 days.

Here's what you get from Nimble that you don't get with Claude's native web fetch:
➡️ Access web content on JS-heavy pages
➡️ Retrieve fresh data instead of cached info
➡️ Build large datasets from the web more efficiently
➡️ Run purpose-built web research workflows

See the difference between Nimble and Claude's web fetch in our benchmark below ⬇️

[Nimble @nimble_search]

Web search for AI Agents is noisy. They can search the web, but most of the HTML they retrieve is useless.

Nimble cleans it up, so your agents get the precise information they need from the web in structured JSON...without the extra junk.

[Ilan Chemla @ilChemla]

@claudeai The "where did I leave that session" tax was real

Three Claude Codes open, five minutes finding the one you actually wanted

Thank you 🙏🙏

[Ilan Chemla @ilChemla]

running 4 Claude Code sessions across 3 repos used to mean 4 terminals and a sticky note

now it's one `claude agents` and the cognitive overhead is just... gone

the agent view ships the productivity gain everyone said sub-agents would. shipped quietly.

[Ilan Chemla @ilChemla]

Agents need to reach the parts of the web that weren’t built for them.

That’s what we work on at @nimble_search.

Good place to start:
https://t.co/aBHouC68KO

[Ilan Chemla @ilChemla]

I had to read the @salesforce Headless 360 announcement twice.

Not because it was unclear.

Because the weirdest line came from a cofounder:

“Why should you ever log into Salesforce again?”

The interesting part is why:

[Ilan Chemla @ilChemla]

No clean #API.
No #MCP tool.
No agent-ready interface.

But agents still need to use it anyway.

That’s why web access keeps getting more interesting to me.

Because:

[Ilan Chemla @ilChemla]

It was never #MCP vs #CLI.

Your #AI Platform needs to support both.

[Ilan Chemla @ilChemla]

People keep asking me what I think about the #MCP vs #CLI debate.

Biggest drama since Tabs vs Spaces.

My honest answer:

we need both.

And no, that’s not me being diplomatic. 👇

[Ilan Chemla @ilChemla]

But here’s the part people skip:

@ChatGPTapp web users are not opening a terminal and running npm install.

Same for browser tools, embedded agents, mobile agents.

That’s why we offer both at @nimble_search.

Different customers, stacks, and agent runtimes.
👇

[Ilan Chemla @ilChemla]

Best part:
👉 @QodoAI reviewed.
👉 The contributor’s agent fixed.

The same file that teaches @claudeai Code how I build now teaches Qodo how I review.

One source of truth.
Two very different jobs.

What’s one rule your coding agent should never ignore?

[Ilan Chemla @ilChemla]

Having an agent review your own PR is cool.

Having it review other contributors’ PRs like you would?

That’s a whole other story.

Last week I built a @claudeai.md for my open source Agent Skills repo.

Not a generic one.

Exact patterns.
Skill structure.
Versioning.
Shared logic.

[Ilan Chemla @ilChemla]

Then I connected @QodoAI to the repo.

Qodo treats CLAUDE.md as a first-class rule source.

It imports the rules, enriches them with scope and severity, and keeps them synced when the file changes.

Then a contributor opened a PR.

Qodo left comments I would have written myself. 👇

[Ilan Chemla @ilChemla]

Treat your agents like a brilliant intern who lost their watch.

[Ilan Chemla @ilChemla]

Agent frameworks are bad at one thing nobody talks about: dates.

Three things every builder learns the hard way:

1. The model doesn't know what day it is.

2. Even when you inject the date, it can drift mid-session. Users got told Monday was Tuesday.

3. Date math isn't math for LLMs. Benchmarks: under 20% accuracy without chain-of-thought.

What I do now 👇

[Ilan Chemla @ilChemla]

Inject datetime + day-of-week at message time, not session start.

Compute deltas in code, not in the model.

Trust the model with reasoning, not with the clock.

Pro Tip: wrap it in a Time skill. The agent learns exactly when to ask the clock instead of guessing.