Articles · 2026

Measurements, not opinions, about the machinery around AI agents.

Two series. One reads the 2026 harness-engineering papers: what the harness costs, how it manages context, whether a model can write its own, what runs on a laptop, and how multi-hour runs hold together. The other reports my own measurements of agent sandboxes and LLM judges. Every number is traced to a table in a cited paper or to a raw trace.

01Harness engineering
September 8, 2026 · 8 min

The 100-line harness and the million-line harness solve the same tasks. What are the other 999,900 lines for?

Open the source of a coding agent called mini-swe-agent expecting the usual sprawl and you find an agent class of about a hundred lines of Python. It has exactly one tool, which is bash. Every action runs in a fresh subprocess, so there...

02Harness engineering
September 8, 2026 · 8 min

The harness change that doubled a local model's fix rate, and the window size where it stops mattering

Anyone who has run a 30-billion-parameter coding model on a laptop has watched the same scene play out. The model reads a file, then reads it again, then proposes a test command that fails, proposes it a second time, proposes it a third...

03Harness engineering
September 8, 2026 · 8 min

Can a model build its own harness? Five papers say yes, for games and math, and not yet for code

The idea is hard to resist once you have heard it. A harness is code, and models write code, so why not let the model write the harness, score the result, and iterate until it improves? Between February and September 2026 at least five...

04Harness engineering
September 8, 2026 · 9 min

Running a coding agent on a laptop in 2026: what works, what breaks, and the numbers I measured today

I spent this afternoon getting two open-source coding harnesses to run a 30-billion-parameter model against real SWE-bench tasks on a MacBook Pro, with nothing leaving the machine. It works. It also took four detours that none of the...

05Harness engineering
September 8, 2026 · 9 min

Three teams built a long-running agent harness this year. They converged on the same three roles.

Anthropic has published two engineering posts about getting an agent to work for hours without a person in the loop. The first, Effective harnesses for long-running agents from November 2025, is about bridging sessions. The second,...

06Sandboxes and evaluation
September 8, 2026 · 9 min

Hardening was free, a stronger boundary removed a protection, and the test that caught it

Every AI coding agent has the same shape underneath. It writes a snippet, runs it, reads the output, and tries again. Dozens of times per task. The sandbox that runs the snippet sits inside that loop, so whatever it costs, you pay on...

07Sandboxes and evaluation
September 8, 2026 · 10 min

Where the time actually goes in an AI agent's sandbox

If you run a coding agent in production, you have a sandbox somewhere, and at some point someone will ask you three questions about it. How much of each agent step does it eat? How many steps per second can one box handle? And what...

08Sandboxes and evaluation
September 8, 2026 · 9 min

Our two LLM judges agreed 0.71 of the time. A regex was doing the agreeing.

If you have shipped an LLM-as-judge pipeline, you have probably shown someone a two-by-two table and a Cohen's kappa. Two judges from different model families, a chance-corrected agreement in the "substantial" band, and the pipeline is...