Sid
Hello there, welcome to my blog. This is a place where I write about my personal interest and experiences in the space of technology, product, marketing, and sometimes philosophy.
Read the blog posts or check below links for more info.
Featured
The Safest Model Is Not the Top Scorer, and My Benchmark Can Prove It
Updated: at 10:00 AMIn August 2026 I built ToolPick, a small open benchmark that asks one question, does the model make the right first tool call when the toolset is full of decoys, policy traps and pressure. The two models tied for the top score included the one that closed unresolved tickets and deleted a file it was asked to tidy. The flagship scored lowest and never broke policy once. Here is how the benchmark works, the numbers with honest error bars, and what I think a team should do with them.
Proof of Mandate, or How to Hijack an Agent and Steal Nothing
Updated: at 04:52 PMThe industry's reflex on agent safety is "make the model safer." I think that is the wrong fight. Assume the model will be fully compromised, and make the theft impossible anyway, by turning the agent's permissions into a keyring it cannot exceed rather than a rulebook it is asked to follow. Here is the design, and a demo where a jailbroken agent tries to drain a treasury and fails at signing time.
When Two People's Agents Negotiate, Here Is What Breaks
Updated: at 10:00 AMIn August and September 2026 my co-founder and I let our personal coding agents coordinate real work through an agent-to-agent engine I built. The agents were mostly fine. The protocol around them broke in specific, instructive ways, and the lessons we wrote down did not survive the next session. The mechanical guards did.
One Signed Quote, Four Protocols, No Price From the Model
Updated: at 10:00 AMI built a kit that lets an AI agent open an account at a Shopify store, negotiate a price and pay, over MCP, UCP, ACP and x402. Every surface serves the same merchant-signed quote, and the price comes from a deterministic merchant policy, never from a model. Here is what I learned implementing all four protocols against one quote, and an honest account of what has not yet touched a real service.
Never Let the Model Do the Maths
Updated: at 10:00 AMI built Mira, an AI servicing agent for loans and credit cards that works over text and live voice. The rule that shaped everything is that the model never produces a number. It decides what to look up and how to explain it; deterministic calculators and a policy rules engine that cites clause IDs produce every figure. Here is how that works, what the latency looks like, and what is still simulated.
How I Built an AI Harness That Ships Products on Its Own
Updated: at 08:34 PMEveryone is waiting for a smarter model. I spent that time building the operational wrapper around the model we already have, a multi-agent harness that takes a one-line task and returns a deployed product for about five cents. Here is how it actually works, and what shipping hundreds of autonomous tasks taught me.
Recent Posts
I Picked Paper Mentor's Writer by Reading It Blind
Updated: at 10:00 AMPaper Mentor writes a self-help book about you, a chapter at a time, from real books found in a public catalogue. I chose its models by reading chapters blind rather than by a leaderboard, and priced it so a subscriber who reads everything still leaves an 80% margin. Then production broke three numbers that every test said were fine, and the app is still in a closed test, not launched.
The Hardest Thing for an Editing Agent Is Leaving Things Alone
Updated: at 10:00 AMClearCut is an agentic video editor I built in July and August 2026. You talk to it, it watches the footage, and it edits a timeline you can see and undo. This post covers the eval that checks whether it knows what not to cut, the prompt caching mistake behind a $2.61 edit, and the three bugs one end to end run found that 181 unit tests could not.
Teaching Claude a Blockchain It Keeps Getting Wrong
Updated: at 10:00 AMAsk a frontier model to build on Supra and you get code that looks exactly right and does not compile, does not deploy, or quietly calls an endpoint that 404s. So we built supra-dev-skills, an open Claude Code plugin that grounds the model in the APIs as they actually are. Here is the catalogue of hallucinations it fixes, how the skills are designed to stop them, and the replayable sign-in bug we found in our own template along the way.
A Paywall for AI Agents, Built on a Forgotten HTTP Code
Updated: at 09:37 AMHTTP 402 "Payment Required" was reserved in 1997 and never shipped. I used it to build a paywall that AI agents can pay through on their own, with on-chain settlement, non-extractable keys, and verification that runs on an edge function with no database. Here is how the handshake works and the decisions that made it safe.