Everything here comes out of building LLM-powered recommendation and agent systems that shipped to real users — and out of what broke afterwards. The through-line: agents are easy to demo and hard to trust, and almost all the engineering that closes that gap lives in memory, evaluation, tooling, and permissions rather than in the model.
01The five
1How Do You Know Which Agent Is Better?
Six agents, one task, one framework for telling them apart — position@K, recall@K, behavioral clustering — with a live simulator that runs all six and shows why outcome-only scoring rewards the wrong one. Start here if you care whether an agent is actually good, which is the question everyone skips.
Evaluation · interactive · ~12 min 2The Last Click
A real open-source contribution, traced end to end: forty minutes to a tested fix, five hours to publication, four permission walls, one human click. A field report with the actual denials and a live trace viewer — and an argument that the walls are the best-designed part of agent infrastructure.
Agent permissions · field report · ~10 min 3The Model Is a Guest. The Harness Is the House.
Why swapping models five times leaves your product intact, and swapping the harness doesn’t. A razor for what survives each migration — model-shape depreciates, world-shape compounds — with a 36-month simulator, and the bitter-lesson counterargument taken seriously rather than waved off.
Strategy · interactive · ~11 min 4Your Agent Answers in Paragraphs. It Should Answer in Interfaces.
Generative UI with a working playground in the page: type a request and watch intent → spec → validation → rendered interface, then use the charts and forms it built. The specs-not-pixels architecture, UI components as tools, and how you evaluate an interface.
Generative UI · live demo · ~11 min 5Your Agent Has a Filing Cabinet. It Needs a Hippocampus.
The four-memory architecture agents are missing — episodic, semantic, procedural, working — and the consolidation loop that turns a transcript into a reusable skill. The piece I’d point at to explain why “just use a vector DB” keeps disappointing people.
Agent memory · interactive · ~12 min02Or follow a thread
The posts are written as five series rather than twenty-two one-offs. If one of these is the problem in front of you, read the series in order:
Building agents that finish
Why long-running agents stall, forget, and repeat themselves — and the memory and loop infrastructure that fixes it.
Marathon → Memory → Hippocampus → Amnesia → No Undo → Loop Infra
Knowing if they’re any good
Evaluation that survives contact with reality: discriminating between agents, catching Goodhart, and deciding how much autonomy to grant.
Long-Horizon Eval → The Wrong Test → Goodhart → Verification Gap → Autonomy Dial
Interfaces: tools, pictures, generated UI
The surfaces agents touch and produce — tool contracts they read, visualizations we need to debug them, interfaces they render.
Tool Design → Seeing Agents → Generative UI → The Last Click
Strategy & economics
Where durable value sits: the harness over the model, domain depth over generality, and what coordination really costs.
Harness House → Domain Agents → Coordination Tax → Agents in 2026
ML foundations
The retrieval and ranking work underneath all of it.
03If you’d rather see code
Every interactive figure on this site is client-side and self-contained — view source on any post and the whole simulator is right there, no build step. Beyond that: GitHub has the open-source projects, the Prompt Studio runs a dual-encoder intent model in the browser, and the research section has the peer-reviewed visualization work from the PhD.
The blog is also gated by its own eval suite: every post is checked in a headless browser on every push — that it loads, that its interactive widget actually boots, that no link is broken — and the build fails if any check fails. It has caught three real regressions so far, which is three more than I expected.
Hiring, or just curious?
I’m open to ML / LLM and recommendation-systems roles. The fastest way to judge the work is to open any post above and use the thing embedded in it — then email sugeerth@gmail.com. I reply within a day. What I’m working on right now is on the now page.