The Boundaries of Agent Autonomy
Autonomy is not a switch from low to high. It is a set of decision rights that must be granted, evaluated, constrained, and revoked independently.
Writing
Notes, decisions, and working theories.
Notes on agents, product judgment, memory, systems, and the practical work of making new technology useful.
2026
08August
13 min
AGENTS.md and lightweight spec-driven collaboration
Long-running collaboration with coding agents needs durable rules and explicit decisions, but not a specification for every change.
Coding Agents as Execution Systems
A technical walk through how a coding agent explores, plans, edits, verifies, and recovers while fixing a bug across UI, API, and persistence.
State ownership in agent products
Reliable agent products begin by deciding which system owns each fact—not by making the reasoning loop more elaborate.
07July
Enterprise AI: tasks, artifacts, and delivery
The value of enterprise AI is not a more fluent conversation. It is turning ambiguous work into outcomes that can be reviewed, executed, and trusted.
Context, memory, and the attention budget
Reliable agent context depends on separating archives, durable memory, and the current working set—not loading everything the system knows.
From Model Calls to Agent Systems
How a model call grows into an agent loop, and why reliable execution requires runtime, context, state, permissions, and verification.
Memory and product continuity
Useful memory is not a larger transcript. It is a product’s ability to preserve continuity without accumulating confusion.
06June
Billing boundaries in generative AI systems
Reliable billing depends on defining exactly when work is accepted, completed, delivered, settled, or reversed—not only on maintaining a price table.
The model attempts behind a successful request
A request can succeed for the user while the infrastructure behind it is already failing. The missing unit of observation is the individual model attempt.
04April
Proactive systems: attention, timing, and feedback
A proactive personal system must know what deserves an interruption, when to stay silent, and how outcomes should change its next judgment.
03March
The lifecycle of a media generation job
Long-running media generation is a distributed job that needs durable state, recovery, settlement, and delivery—not a single API call.
01January
Notes on agent runtime design
Native tool calling made Oracle's decisions cleaner, but reliability still depended on state, inheritance, verification, and recovery.
2025
11November
Squad: a postmortem
Squad could run and replan for months. It still could not evaluate whether today's action would remain correct over a long horizon.
08August
State and recovery in long-running agents
Model inference happens inside a call, but real work unfolds over time. A reliable agent must pause, recover, remain visible, and turn completion into delivery.
05May
Oracle: from answers to tasks
When a user gives AI a goal rather than a question, the fundamental unit of the product must shift from the answer to the task.
2024
12December
How to Actually Give a Web Page to AI
A URL is not content, and copy-paste is not the answer. A web page has to travel through an easily overlooked path before AI can use it as material.