Series on Guy Freeman

Disciplining the Coding Agent

Wed, 26 Aug 2026 00:00:00 +0000

Four posts on the same problem from four angles. If you did not write the code and cannot read all of it, what is left to trust?

The answer each post converges on is: not the code, and not the model’s account of itself, but the environment — a contract the agent cannot renegotiate, a decision rule that was allowed to say no, an answer key sealed before the first line was written, and a standing assumption that every one of those guards is dead until something has tried to kill it.

Credence

Mon, 03 Aug 2026 00:00:00 +0000

Credence treated which tool to query as a decision problem under uncertainty rather than a language-generation problem. These posts follow it from the first claim to the last measurement.

They are not all in agreement with each other, and that is deliberate. Several of the early posts assert a margin — the Bayesian agent beating an LLM agent by roughly 120 points — that a fairer benchmark later reversed. Rather than quietly edit those posts, each carries a correction note pointing here. If you read only one, read the retrospective, which sets out what the benchmark was actually measuring, what survived, and what did not.

proplang

Sun, 26 Jul 2026 00:00:00 +0000

The design premise is that a rule you can break is not a rule. proplang makes the wrong program unsayable: there is no syntax for forgetting, no syntax for an unbounded deliberation, no syntax for an agent that prefers to be right at any cost.

Each post takes one thing the language will not let you write and argues that the absence is the feature. The build discipline behind it — tests frozen and signed before any code existed — is covered separately in Disciplining the Coding Agent.

The Tool-Call Governor

Fri, 19 Jun 2026 00:00:00 +0000

A coding agent makes tool calls you did not authorise and re-runs ones it already made. Both are decision problems, and both are cheap to get right if you are willing to hold a posterior over what the agent is doing.

Four posts: the body–brain split and why the brain must stay opaque to the body; what a regex cannot do that a calibrated governor can; and two releases of the resulting OpenClaw plugin — the second of which asks you to reproduce its numbers yourself rather than take them from me.

The Velotix Files

Wed, 03 Jun 2026 00:00:00 +0000

An investigation, in five parts, built on government enforcement findings, court filings, archived pages and public records.

Disclosure: I am a former Velotix employee (October 2024 – May 2025), and I filed the Ministry of Labour complaint referenced throughout. Where I draw on my own experience rather than the documentary record, I say so.

Part 4 documents a fraudulent DMCA notice filed against Part 1 from an address in Chittagong, citing an unrelated Israeli newspaper article. Part 5 covers the company’s recovery, as told to Maariv.

Agents from First Principles

Tue, 28 Apr 2026 00:00:00 +0000

The question underneath all five: what would an agent look like if you derived its behaviour from a few fundamentals rather than assembling it from components?

It starts with a creature that has to learn which things are safe to eat, and ends with a claim that the entire apparatus is three types and four axioms. Along the way it plays Zork badly, then less badly, and the reason it improves turns out to have nothing to do with intelligence and everything to do with how it represents where it is.