Mac Anderson

Founder and CEO of Oxagen

I build coding agents that ship software on their own.

I research how coding agents work and where they fail. I build them too, including agents that write their own tools. The goal is software that agents deliver on their own, and that still scales and stays easy to maintain. I wrote code by hand for 16 years before I handed all of it to agents in 2026.

27,022GitHub contributions since 26 Feb 202658 of 89Stella on Terminal-Bench 2.121 partsField manual, free to read

Activity

27,022 contributions in 225 days

Each dot is one day on GitHub. Before 26 February 2026, Mac wrote his code by hand. After it, agents wrote the code and he directed them, read their work, and decided what merged.

Updated October 9, 2026Hover or tap a dot to see the day
Contributions in all of 2025
38Before agents wrote the code
Contributions since 26 Feb 2026
27,022225 days
Merged pull requests
7,102All time, every repository
Most in one day
1,144August 30, 2026
Longest daily streak
62Days in a row
Active days in the last year
211Out of 365
How these numbers are counted
  • Contributions are GitHub's contribution calendar counts, which include private contributions Mac has chosen to show.
  • This total adds up every day from February 26, 2026, when agents began writing Mac's code, through today.
  • Pull request and issue counts are all-time GitHub search results, which cover only the repositories macanderson's token can see.
  • The longest streak is the most days in a row with at least one contribution each.

Field manual

Engineering Deterministic AI Coding Agents

A free field manual in 21 parts. Parts 1 to 13 show how to build the system around a coding agent so it costs less and behaves the same way each time. Parts 14 to 20 show how to run the agents you deploy. Every part cites its evidence and ends with steps for this week.

parts
21 parts
cited sources
27 cited sources
checklists
21 checklists
hours to read it all
2 hours to read it all
Start reading Take the self-assessment PDF

The economics

  1. 1Agents are not expensive. Bad architecture is

Deterministic retrieval

  1. 2Stop making agents read your logs
  2. 3Parse the prompt before the model sees it
  3. 4Why grep is the wrong retrieval engine
  4. 5Build semantic memory once

Context engineering

  1. 6Rules belong in data, not in prose
  2. 7Compression beats a bigger model
  3. 8Show the agent the schema

Orchestration and memory

  1. 9Why coordinator agents do not scale
  2. 10Tests as retrieval objects
  3. 11Working memory, not bigger windows
  4. 12Knowledge graphs over concatenation

Measurement

  1. 13Measuring an agent in production

Operating the workforce

  1. 14Give every agent its own identity
  2. 15Write the mandate
  3. 16The agent asks, a rule decides
  4. 17Spend you can attribute
  5. 18Equip agents on purpose
  6. 19Keep a record another person can read
  7. 20Bounded tasks and ongoing work

The endgame

  1. 21Better systems, not better models

Research

Latest research

All research
  1. October 9, 2026Building a continuous delivery system for fine-tuned open-source, open-weight models trained on your organization's tracesEvery coding-agent session your team runs produces a trace. Kept and graded by an oracle the agent cannot touch, those traces become the training set for a model you own. This book covers the oracle, the air gap, the one-bit verdict, the harness hooks that collect traces for free, how much data a fine-tune needs, and the pipeline that delivers new weights every week.109 min
  2. September 30, 2026Agents are waiting on a process built for peopleAn agent can write a change in minutes. Then the change waits for a person to read it. This post measures that wait, what eight companies changed about it, and how Oxagen ships at every hour.11 min
  3. September 16, 2026Steering a run you are not watchingWhen a long run goes wrong, most teams can stop it or type at it. Both work badly. A steer is a third option. It is a message with a delivery mode, a status, and a record.7 min
  4. September 16, 2026The agent time horizon is doublingMETR measures how long a task an agent can finish on its own. That length has doubled about every seven months since 2019. This post covers what week-long runs mean for supervision.7 min
  5. September 16, 2026The problem is not slop, it is your processThe worry about AI slop is about output quality. The measurements point to a different cause. Teams give an agent a workflow built for people and expect it to work.7 min

Work

Products

Company

Oxagen

Workforce management for autonomous agents. Each agent gets an identity, a mandate that sets what it may do and spend, the tools it needs, and a record of what it did.

Wraps Claude Code and Codex

Open source, Rust

Stella

A fast, model-agnostic terminal coding agent. Its definition of done is a witness test: one that fails on the old code and passes on the new.

58 of 89 on Terminal-Bench 2.1, against 44 for Claude Code on the same model

Benchmarks

Arena

Runs a coding agent head to head against Claude Code, Gemini, and other agents, with regression checks that keep it improving instead of drifting.

Rules published before each run

Open specification

Context Graph Protocol

A typed graph format for an agent's context: how to represent it, exchange it, and account for every token of it.

SDKs for TypeScript, Python, and Go

Mac Anderson, black and white portrait

Story

Sixteen years of code by hand

“verification can not lie”
Mac, in a prompt on 31 July 2026

In 2026 Mac stopped writing code by hand. Between 28 May and 20 September his agents made 6,341 commits across 13 repositories, and he learned that the hard part is not prompting. It is knowing what each agent did, what it cost, and who allowed it.

Read the story

For agents

Agent access

This site is written for AI agents as well as people. Point yours at it and it reads the same facts you do.

See the site blueprint
  • /llms.txt

    A map of the whole site in the llms.txt format.

  • /llms-full.txt

    Every page in one markdown file.

  • /manual/write-the-mandate.md

    Any page as markdown. Add .md to the address.

  • /mcp

    A read-only MCP server with search and page tools.

  • Add the MCP server to Claude Code:

    claude mcp add --transport http macanderson https://macanderson.com/mcp

Subscribe

Updates by email

New research, field manual updates, and the first episodes of Mac's video series go to subscribers first.

No spam. Unsubscribe any time. Read the privacy note.