Spent more time learning deepagents from @LangChain with primary focus on integrating MCP servers with auth that doesn't conform completely to the OAuth standard. ( In prep of a work coming my way ) Below POC is on top of yfinance api by encapsulating it in an MCP server. MCP server isn't a 1:1 api mapping with yfinance, but a set of meta api with focus on valuation, quick ratios, fundamental analysis and clubbing gemini web search for grounding before responding ( not recommended as you can spot hallucinated or wrong information being picked up on pricing ). Inherited OAuthProvider libs/code/deepagents_code/mcp_providers/base.py and made the changes for custom provider. Found deepagents codebase easy to understand and extend. If you are looking to build a custom harness for a very domain specific task where you want to bring your own compaction logic, enforce token quotas / max tool calls to make and enforce some sequence based on rule engine / switch tool selection logic itself to an GEPA optimized prompt / intercept tool calls and pass them thru a custom guardrail before and after call it's easy to find the interface, understand the functionality and swap it. It's thoughtfully designed such that you can swap sub-components and bring just want you need. Let million harness bloom.

Morning Market Grocery Run With Son in Ujjain - MP.

Getting LLMs to dumb down and explain medical and blood test reports is honestly the no 1 god send feature. Plus adding the report and medication and asking why style of questions is helpful to educate. Like what should be the questions i should ask the specialist and why?. Don't distrust your dr but always verify and ask questions in good tone and respectfully i guess. Plus for chronic life long health issues that are medically managed you can quiz on progression and then to manage it. If anything you know what's coming medium to long term.

15 LLMs predict FIFA world cup scorelines based on web search, news, players data, and Polymarket odds! Built with DSPy! Website and code in comments.

At scale with models in context of tool calling here are the failures you are bound to see - Wrong parameter in invoke - Missing parameter that's mandatory - Wrong tool name entirely sent in invoke - Random Malformed tool error when it isn't the case and just works fine with retry - Spec Format Drift from LLM when doing invoke - Type Confusions - Calling same tool again and again despite getting successful response in previous invoke Larger frontier models can self correct once you send error, telemetry/trace, stack trace etc with retry as followup in next turn. Smaller models not so sure.

enabling rubric middleware is one of the easiest ways to add "loops" to your agents this loop makes the agent keep doing until acceptance criteria are all met!

Most tool-calling sandboxes projects off GitHub will require privileged access, CAP_SYS_ADMIN, and user namespaces and cross platform quirks that a team where devs have both windows and Mac, and deployment is linux. During local dev Qwen3.5-0.8B/2B with WASM/Pyodide Deno ( extracted from DSPy ) are 0 cost combination to plug in with no cross platform quirks. Necessity is the mother of all inventions, Lack of Funds is the mother of all resourcefulness.

Huge thanks to the @AnthropicAI team for investigating `just-bash` with Mythos as part of Project Glasswing There are no serious findings. I'll issue a release with minor hardening improvements after my vacation. This is an interesting case for 2 reasons: 1. `just-bash` is a very ripe surface (it's a full implementation of bash and common utilities, bundles QuickJS, CPython, has optional filesystem access) 2. `just-bash` itself was largely written by Opus 4.5 with minimal human review (but deep hardening loops and very paranoid machine-enforced coding rules) justbash.dev

My X feed right now

Reaffirmation in long email is how you know LLM wrote it and the person sending it didn't read it.

a reminder that branding matters.

AI isn't software you buy once - it's a raw commodity we are currently renting via foreign APIs. If this continues, our AI lease bill in five years could eclipse what we spend importing crude oil or gold, trapping us in a massive digital extraction loop. indiatoday.in/india-today-in…

another vid from the goat of beautiful animated explaination even if you understand entropy and its derivations through and through, this is still interesting refamiliarize with some really nice examples youtu.be/l6DKRf-fAAM?si=NYiv…

Worthy watch youtu.be/PJ-3hXAUotI?si=VOk9…

By virtue of having a kid much earlier than our peer age group we got confronted with certain experiences/challenges before most. Over the years we have found it super helpful to ask those whose kids are like 2-8 years older than our kid's age for advice on certain decision making or how they navigated things. Most of times the response is mostly on the lines of - we made this mistake plz don't make it 😅 so do this instead Here are few we keep getting to hear - Force kid to commit to one hobby or sport for a longer duration like couple of years, don't allow them to quit and try something new every year - If you are going to take your kid in a car everywhere all the time then once they hit college or a job prepare to buy them one ( this is a personal decision i think based on your financial standing ) - Send your kids to live for few weeks either in summer camp or at relatives place alone before their first experience living alone in a hostel - Teach your kid how to cook quick 4-5 different breakfast and 2-3 meal items and have them do it multiple times before 12th grade so they don't struggle during their first job if living alone in a new city Parenting is like old classic video games, just when you start getting comfortable and a hang of things the level changes and you are in a new terrain / level. Looking to reach the zen level of my parents who now say, ohh we didn't overthink like you guys, everything takes care of itself.

OpenClaw hit 3,000 commits in a single day. 10 to 15 maintainers. All with day jobs. @vincent_koc (Chief Architect of OpenClaw) explains how the factory actually works. youtube.com/watch?v=pmoDeA3R… The great refactor: 2 AM, Vincent and Peter at NVIDIA, 60 to 70 agents running between them. 2,700 commits. Close to a million lines changed. 82% of the core codebase touched. Plugin architecture shipped by morning. The saving grace: overfitted unit tests AI code loves to generate. As long as they went green, they knew they were close. Knowing when your agent is bullshitting you is the skill nobody talks about. "It doesn't sound off because of what it's doing. It sounds off because of how it's explaining itself to me. It's waffling." 2025 was about token maxing. 2026 is about not wasting them.

Been learning how to get better at image generation, sort of becoming a weekend 1-2 hr hobby now. Seed Prompt -> Brainstormed using GEPA to enrich seed prompt with more details Image Reference ( for below image reference came from the work done by the brilliant Vigo Zhao @VigoCreativeAI - Go follow him ) Next thing am planning to do is build a small image subsection selection tool so that I put references on what to fix on the image itself and pass it to image models. So easy to get typo fixed and reducing misunderstanding by model on what change has been asked.