Series
Most of what I write here belongs to a longer argument. These are the sequences — each one read best in order, though every part stands on its own.
Disciplining the Coding Agent
When a model writes the code you cannot supervise the lines, so you constrain the environment instead: a spec as contract, an adoption gate with the power to refuse, acceptance tests frozen and signed before any implementation existed, and finally an agent paid to break all of it.
Credence
A Bayesian decision-theoretic framework for LLM agent tool selection: what it claimed, how it was measured, what shipped, and the one-line ablation that eventually beat it. Seven posts, ending in the retrospective that retracts the headline.
- Agentic AI Is Neither Intelligent Nor an Agent
- How Decision Theory Cuts Your AI Agent's API Bill in Half
- Sixty-Two Percent Correct and Winning by a Hundred and Twenty Points
- The Prompting Gradient
- The Agent That Invents Its Own Rules
- Ninety-Six Percent Cheaper and Slightly Better
- The Ablation That Beat the Agent
proplang
A small language in which a Bayesian decision-theoretic agent is a program, cornered so tightly that the architecture is enforced by the grammar rather than by good intentions. Five posts on what it refuses to let you say, and why each refusal is load-bearing.
The Tool-Call Governor
credence-pi: a Bayesian governor that sits between a coding agent and its tools, blocks the calls it wastes, asks before an injected action, and routes each turn to the cheapest model that can handle it. Four posts, from architecture to shipped plugin.
The Velotix Files
A five-part investigation into Velotix Ltd, an Israeli cybersecurity startup that marketed angel investments through an ISA-licensed crowdfunding platform while under government enforcement for not paying its employees — and what happened after Part 1 was published.
Agents from First Principles
Five posts building a Bayesian learning agent from nothing: a creature that learns what to eat, an evolutionary sequel, a text adventure played through a value-of-information gate, the state-representation fix behind it, and the three types the whole thing reduces to.