ep 06 field notes
Show Us Your Agent Skills / EP 06 / guest dossier
MATT ROCKLIN DASK COILED FRISKY AGENTS.MD

MATT ROCKLIN

Matt runs Coiled in five to ten hours a week, largely with agents that read a directory describing the whole company: accounting, engineering, customers, legal, insurance. Those agents found customers with contracts and real usage who were never being billed. His Rust projects run the same way: Frisky, a Dask-like distributed system, exposes dashboards, telemetry, and a documented CLI so agents can inspect the system they are changing. He writes the plans by hand, walks away for an hour or two, and reviews with targeted evidence: benchmarks, algorithm explanations, line counts, fresh agents.

▶ WATCH SEGMENT $ COMPANY-CONTEXT $ FEEDBACK-SYSTEMS
5-10 HRSa week to run a company making one to two million a year BILLINGagents found contracted customers using Coiled who were never billed 10Xconcurrent agent loops; "my brain is still kind of Swiss cheese" NO BULLSHIThis least-boring reason to work with agents 1 WEEKfor an agent to rebuild NumPy in Rust against its own test suite JANUARYhe has barely read code since
EP 06 · MATT ROCKLIN · broad company context, feedback systems, and long agent turns, live on stream

"People should drink more when writing agents."

Architecture writing is best done in the morning or after drinks. Agents make it easier to act on strange, social, half-formed ideas before the old sense of software difficulty shuts them down. 00:10:28

THE WHOLE COMPANY IS CONTEXT

the company-context-agents workflow

Coiled used to be a VC-funded startup. It now runs as what Matt calls a lifestyle company: five to ten hours of his week for a business making roughly one or two million dollars a year, with agents doing much of the operating.

The operating surface is a directory that describes the company, with sections for accounting, engineering, customers, legal, and insurance. Matt says the agents "actually understand the full company in a way that historically no individual has," and they are "B plus programmers that are also B plus lawyers." That reach found customers with contracts and heavy usage who were never being billed, and it rewrote Coiled's master services agreement into a document far more tuned to the company.

The same discipline runs at project scale. His AGENTS.md files reach five to ten thousand words, all handwritten, culled and refactored every week or two. They carry principles rather than how-tos, so every agent starts with a hundred lines on the concepts that recur through the repository.

Matt Rocklin explaining how broad company context lets agents reason across departments
The Coiled context directory on stream: agents connect contracts, usage, accounting, product, and legal in one pass. [00:14:23]

"All of it."

Skylar asks how much of a five-to-ten-thousand-word AGENTS.md file Matt writes by hand. This from a man who has barely read code since January. The reading stopped; the writing moved into the instructions. 00:17:38

PLAN THE TURN, WALK AWAY

the agent-feedback-systems workflow: evidence first, long turns, review by concern
Decide what evidence countsBefore any work starts, figure out what will show whether the job is being done well. 00:22:51
Write the plan by handHe sits, thinks, and writes a markdown file in Vim, then hands it to the agent. 00:09:20
Walk away for an hour or twoProductivity continues without him watching. He runs the loop ten times concurrently, and three dumb runs out of ten are fine. 00:23:01
Encode repeated worries as feedbackWhen the same review question keeps coming up, like "is this worth the complexity it added?", he turns it into a question the agent always gets. 00:19:45
Review by concernComplexity gets an algorithm explanation, speed gets benchmark comparisons, redundancy gets a check that the new system is distinct. 00:38:33
Finish with a fresh agentA reviewer with clean context is the standard closing step. 00:23:11

"AI observability often means observing your agents rather than giving your agents observability over the system that they're building."

Profilers, test suites, telemetry, and a documented CLI give the agent the system's internals while it builds. 00:31:22

OBSERVABILITY FOR THE AGENT

Frisky, Matt's Rust replacement for Dask, and the Frisky and Xarray example

Frisky is Matt's Rust replacement for Dask, the distributed Python framework he created. It started as an art project, then came back during consulting work on large financial services datasets: "I was using Dask and Dask was annoying to use because it was both slow and opaque." Rust made a ton of telemetry cheap enough to leave on, so the dashboards and internal state exist for the agents as much as for him.

Frisky ships a documented CLI. Agents follow the CLI pages, run more commands when they need detail, and get feedback about what is happening inside the distributed system. When they see the telemetry, Matt says, "they're like, oh, this is delightful."

His first AI-built product was a Rust reimplementation of NumPy that converged in roughly a week because the NumPy test suite and benchmarks were right there. The code stays small too: Frisky lands at about the same line count as Dask, and the NumPy rebuild at about a tenth of NumPy's.

Matt Rocklin showing the Frisky dashboard during Episode 6
Frisky's dashboard and telemetry, one of the feedback systems agents can inspect while changing a distributed system. [00:28:53]

"Even a fairly dumb agent can build something very robust if they've got access to very good feedback."

The NumPy-in-Rust rebuild converged in roughly a week because the test suite and benchmarks supplied the feedback. 00:31:43

THE OPERATING KIT

two workflows, a reconstructed AGENTS.md, and the posts behind the segment
workflow

company-context-agents

Give agents a directory that describes the whole company. The billing gap and the MSA rewrite both came from that one surface.

workflow

agent-feedback-systems

Plan around the feedback signal, leave the agent alone for long turns, and encode recurring review questions so every agent gets them.

artifact

AGENTS-reconstructed.md

Frisky's AGENTS.md rebuilt from the visible parts of the screen share: project overview, development tools, usage, and a performance section.

practice

fresh-agent review

"People know to ask for a fresh agent to come in and review things." A reviewer with clean context closes out the work.

"Agents can be a very dehumanizing experience. They can turn you into machines that give permission."

He resists by watching how he thinks during agent work and encoding the better patterns. 00:19:12

MATT'S STACK

what runs the company and the code
Codexhis main agent; prose that costs less attention to read Claudestill in rotation; million-token contexts remain workable Coiledthe company the agents help run Daskthe reference system Frisky reworks in Rust Rustmade heavy telemetry guilt-free NumPyits test suite drove the week-long rebuild Vimwhere the markdown plans get written Pilocal models for fast learning loops

"The value proposition of Python is low."

From the creator of Dask, after years in the Python data ecosystem: if agents absorb the language-understanding burden, the old ergonomics argument changes. 01:26:15