A small team of researchers-who-build. We ship agent systems that survive Tuesday mornings — eval-first, MCP-native, runtime-first. Skills that get reused, evals you can re-run, agents that don't fall off the 50-step cliff. And sometimes we chase a paper down a rabbit hole purely for the sport.
Half our time is client work — agents, evals, post-training, teaching. The other half is whatever the group chat can't stop talking about this month. Same people, same monorepo; each half makes the other one better.
We design the runtime your agents live in — tools, intake, memory, gates, escalation paths — purpose-built for your industry, not adapted from a demo. Bio labs, legal, dispatch so far. MCP-native, sandboxed by default, boring on purpose. If your last agent nose-dived at step 40, we know that cliff personally.
2–4 weeks · flat fee · you keep the repo
If you're deploying an agent and want to know whether it works on Tuesdays — not in demos — we build the harness: process-aware, contamination-checked, no cherry-picked seeds. You get a reproducible suite that runs in your CI, not a PDF.
fixed scope · your data stays yours · repo + report
RLVR, process rewards, on-policy distillation, synthetic-data pipelines, Skills that hold up under adversarial use. We'll also tell you when a better harness beats a fine-tune — it does more often than anyone admits.
4–8 weeks · open-weight friendly · model-agnostic
Workshops for teams ready to graduate from chat tabs to CLI agents with CI-gated autonomy. Hands-on, no slides — we bring the harness, you bring your gnarliest workflow.
1–3 days on-site · or 4 weeks async
Small experiments nobody's paying for yet — skill libraries that improve themselves from their own traces, runtime patterns that transfer between verticals (and the ones that never will), agent memory that outlives the context window. Published when it replicates.
ongoing · notes below · repos on GitHub
We build tools for ourselves and publish the ones that survive a month of daily use. A few live in the Claude marketplace now. MIT by default — stars appreciated, issues appreciated more.
open source · skills + mcp · PRs welcome
Weekly; attendance optional, opinions mandatory. We read the paper so you don't have to — and sometimes our writeup outperforms the baseline.
public archive · posts in Notes
The one-weekend kind. Computer-use agents speedrunning our own expense reports, voice loops that answer before you finish asking, Skills that absolutely should not ship. Every so often one grows up into a real project.
irregular · no SLA · no regrets
We don't do discovery phases or steering committees. We do a call, then a prototype, then a demo. Most of our engagements look roughly like this:
You describe the problem. We ask annoyingly specific questions. By the end we either send a scope or politely decline.
Shared repo from day one. Traces land in your observability stack, not ours. Friday demo on the real thing, never a staged one. No status decks — the commit history is the status report.
We build the eval suite before we build the polish. You keep the code, the eval harness, the traces. We stay on-call for six more weeks.
We'd love to show you a polished case-study carousel. We also love our clients' confidentiality more. Here's what we can say — ask us about the rest on a call and we'll show you what we're allowed to.
Each line is a real engagement. Names and domains redacted; numbers and methods real. If one sounds like the thing you need, that's a good reason to get in touch.
We met in university. Most of us are still there — some studying, some already shipping, all still meeting in the same group chat. If we work with you, you'll meet everyone who's relevant.
Hi, you found our lab notebook. We publish it as-is: dates, code, traces, and the "well, that didn't work" parts — those are the useful ones. In spirit it's closer to an Obsidian Bases view than to a blog — more #research channel than Medium. If a note saves you one dead end, it did its job.
Tell us in two paragraphs what's going on and what "done" looks like. If it's a fit, we'll send back a scope within a week. If it isn't, we'll say so, and probably point you to someone better.