Syed Mohammed Faham

Trying to make sense of AI

[Syed Mohammed Faham @iamfaham]

Most AI tools are single player: one person, one private chat, one agent nobody else can join.

Multiplayer AI flips that: teams and agents working together in shared sessions, not private bubbles.

YC's Fall 2026 RFS called it out as a category to build in.

[Syed Mohammed Faham @iamfaham]

@anupamrjp https://t.co/FOV67JW8xZ

[Syed Mohammed Faham @iamfaham]

@lightsilver323 https://t.co/7nL9ucJM29

[Syed Mohammed Faham @iamfaham]

pip install agentsnap

pytest plugin + a small CLI.

Repo → https://t.co/z71W9USjH7
Docs → https://t.co/X5UOBCs4UD

Would genuinely love feedback on what breaks on your agents.

5/5

[Syed Mohammed Faham @iamfaham]

Your AI agent regressed and nothing failed.

No exception.
No red CI.
It just quietly got worse and you found out in prod.

I built agentsnap to catch that. Snapshot testing for agents: record a golden run, fail the test when behavior drifts. 🧵

[Syed Mohammed Faham @iamfaham]

Zero instrumentation.
It patches the SDK classes, so raw openai/anthropic (sync + async), streaming, the Responses API, and LangChain / Pydantic AI / OpenAI Agents SDK all just work — no wrappers.Framework support is verified in CI against the real libs.

4/5

[Syed Mohammed Faham @iamfaham]

Two modes:
🔁 Replay — recorded responses fed back to your agent. No API calls, deterministic, free. Run it on every PR.
🌙 Live — real calls vs the current model, catches drift. Run it nightly. The exact pattern: replay on PRs, live nightly.

3/5

[Syed Mohammed Faham @iamfaham]

It records your agent's LLM + tool calls once and commits a snapshot — like Jest snapshots, but for agent behavior.Then it checks 4 things on every run: tool sequence, arguments, which tool the model chose, and the semantic content of responses.

2/5

[Syed Mohammed Faham @iamfaham]

A dashcam. A free API. Public GIS data. Open-source Python.

That's the entire stack behind an AI system that automates housing code enforcement — for $0/month.

Full breakdown ↓
https://t.co/Clpsaw5Dm7

[Syed Mohammed Faham @iamfaham]

The AI model is not your bottleneck

If you're building agents, the bottleneck is the scaffolding around it

Execution, evals, observability, safety, memory - that's harness engineering, and it's where the real work happens

wrote a full breakdown: https://t.co/3FILzj0TLN

[Machina @EXM7777]

here's how to get genuinely good at agentic coding:

Anthropic published a 33-page official guide on Claude skills... everyone skimmed it or ignored it

here's what to do instead:

send it to Claude and have it build you a full course around it

then while you're building your project:
> at every milestone, tell Claude to document what it did and how in a markdown file
> every bug you hit, don't patch it and move on... make Claude trace the root cause using first principles, fix it, then document the whole breakdown

when the project is done, take all that documentation and feed it into skill-creator

congrats, you just trained a reusable skill for this specific architecture (for any SDK, MCP, API...)

next time you build something similar you'll literally one-shot your project

just start stacking skills, it might just change your life

[Machina @EXM7777]

build the habit of reading Claude Code updates, you will unlock new possibilities without having to install a single plugin/extension

[Machina @EXM7777]

if you can't get any work done because Claude is down, you're literally a slave...

you number one priority should be building model-agnostic systems, having a structured knowledge base and system prompts you can swap from one model to another

relying on a single provider is VERY dangerous for your business

[Syed Mohammed Faham @iamfaham]

Great read for anyone curious about the limits of LLMs in numerical forecasting: https://t.co/YjodXCl37X

[Syed Mohammed Faham @iamfaham]

Does adding "context" via LLM embeddings actually make for better forecasts?

@TeachTheMachine put it to the test using DJIA data and news headlines.
The result: A slight edge in accuracy, but no silver bullet.

#AI #DataScience #MachineLearning #Python

[Syed Mohammed Faham @iamfaham]

Two ways to implement:
- Declarative: HTML form attributes
- Imperative: JavaScript tool registration

Tools inherit user auth. No backend server needed. Collaborative browsing, not headless automation.

Docs: https://t.co/tiMezYJnu5

[Syed Mohammed Faham @iamfaham]

WebMCP is now in Chrome 146 Canary. It turns websites into agent-ready tools via the `navigator.modelContext` API.

Instead of scraping HTML, AI agents call structured functions directly. Result: 67% less compute than visual interactions.

[Syed Mohammed Faham @iamfaham]

I break down the idea + a simple experiment in my new article:
https://t.co/SFLST24c4j

[Syed Mohammed Faham @iamfaham]

Most folks think LLM “control” = prompt engineering.
But prompts only steer input context.

What if you could steer internal thinking?

That’s activation steering — shifting hidden states to bias behavior w/o retraining.

[Syed Mohammed Faham @iamfaham]

@Paytmcare When can I expect a response, details have been shared weeks ago on DM. @Paytmcare @Paytm

[Syed Mohammed Faham @iamfaham]

@Paytm @Paytmcare My flight refund request has been pending for 3 weeks despite having written confirmation from @airindia itself. The in-app Help & Support has not been helpfu at all. Please look into this urgently. #PaytmSupport #FlightRefund