Agents

Essays on AI agents and coding agents: what harnesses, tools, memory and multi-agent setups actually do when you read the logs.

Thirty-Six Percent More Pull Requests, Fifty-Six Percent Longer to Merge

18 minute read

· Agents

A Harvard job-market paper built on Jellyfish data from 655 firms finds that a year after adopting coding agents, firms opened 35.7 percent more pull requests per engineer, merged 26.0 percent more, and took 56.1 percent longer to get each one from submission to merge. Resolved Jira issues rose 13.7 percent, which is not significant. The paper’s own numbers imply review inventory roughly doubled, and the code-review literature explains both why the wait grows and why a mean of 3.56 days is a weak instrument.

71% Fewer Tokens in One Harness, 14.5% in Another

13 minute read

· Agents

A paper making the rounds says a reading layer called SparseRead cuts agent tokens by up to 92.9% and keeps quality in 97% of settings. I read both versions of it. The design is sensible and the paper is unusually open about where it loses. But the headline number comes from a model whose default agent was failing the task, the same model saves 71.3% in one harness and 14.5% in another, and the 97% depends on which scenarios made it into the main table.

Ninety Percent Was the Minimum for a Decent Rating

12 minute read

· Agents

A former Meta engineer writes that generating at least 90% of his code with AI was the minimum for a decent rating, and that AI-heavy codebases end up with more incidents that take longer to fix. The controlled evidence I could find measures smaller units than that: one function, one task, one developer. Those mostly look fine. The one large dataset that measures delivery found stability falling as AI adoption rose. The claim about incidents is the one nobody has tested properly, and the 90% target measures how much code AI wrote, which isn’t the thing that breaks.

Forty-Six of Forty-Nine Agents Wrote to Ada First

16 minute read

· Agents

YC Labs and Olam Labs put 50 GPT-6.1 Sol agents in a Minecraft world and then ran Claude-versus-GPT swarms in Bed Wars and Mindustry. The coverage went with GPT beating Claude and a colony that drowned in bureaucracy. The finding that held up best was a different one: 46 of the 49 other agents sent their first message to whoever was first on the roster, and she ran the colony for three hours.

Sixty of Sixty-Six Cost Ranges Overlap. The JavaScript Bytes Barely Do.

14 minute read

· Agents

Nolan Lawson had Opus 5.5 build the same PokΓ©mon app five times in each of twelve JavaScript frameworks and published the medians and the ranges. If you read the ranges, cost, time and tokens barely separate the frameworks: 60 of the 66 pairs of cost ranges overlap. JavaScript bytes sent separate almost every pair. What a framework ships is a property of the framework, and what the agent spends building it is mostly the agent’s luck.

Nineteen of Eighty-Three: The Kernel Passes, and Still Changes Its Answer

15 minute read

· Agents

A paper posted to arXiv on 27 October audits the GPU kernels that kernel-writing agents have already published, and finds that 19 of the 83 with a real batch dimension give a request a different answer depending on who else is in the batch. All of them passed their correctness checks. The fix costs about 8 percent of the speedup.

AI Agents Finished 167 of 400 Psychology Experiments. Their Data Looked Human on the Biases, Not the Abilities.

7 minute read

· Agents

CogArena sends AI agents through ten online psychology experiments in a real browser. Four economy-tier models finished 167 of 400 sessions; most failures were the agent declaring itself done early. Where the data could be scored, agents reproduced human bias signatures 66% of the time and human capacity signatures 31% of the time, which matters for anyone worried about agents slipping into online studies.

The Score Survived Compaction. About One Rule in Three Did Not.

19 minute read

· Agents

A TΓΌbingen group ran four frontier models through OpenCode on three long-horizon benchmarks and found that when the agent compacts its context barely matters to the score. The same summaries drop about one of three anti-cheating rules the first time they are written and about two by the tenth, and on real coding sessions they lose instructions about how to work six times as often as instructions about what to build. The benchmark can’t see that, because everything the score needs is on disk.

Same Weights, Different Harness, Up to 11.5 Points

10 minute read

· Agents

Google Research has released HarnessΒ², an Apache 2.0 framework that lets an agent rewrite its own prompts, tools, memory and skills while the model stays frozen. The paper’s gains are real and largest where the starting harness was weakest, and the judging setup deserves a close read.

83% Byte-Exact, and the Script the Agents Tried to Edit

10 minute read

· Agents

A write-up about agents decompiling a first-person shooter says they repeatedly tried to change the checker that graded them. It also leaves a headline of 500+ billion tokens beside a body estimate of 600 to 700 billion, and a headline pass rate that counts functions.

The Skills Helped in All Nine Setups, Instruction Included. The 36.78 Percent Counts Something Else.

15 minute read

· Agents

A new study of agent skills is circulating as β€˜skills hurt on 36.78% of tasks.’ Its own table says the skills, together with an instruction to use them, raised the average pass rate in every one of nine model-harness configurations. The 36.78 percent counts tasks where configurations disagreed, measured on three runs each, which is a number that noise alone can fill for tasks near a coin flip.

Three Spend Figures and One Oracle

9 minute read

· Agents

A Rust port of the TypeScript compiler says an LLM wrote it, that its compatibility is 100% on every real project it tested, and what it cost. The README gives three dollar figures that do not sit neatly together, and a compatibility line that its own status section qualifies.

Twenty Tries, Two Numbers

17 minute read

· Agents

Microsoft’s ThinkingBox grades agents on the database they leave behind and repeats every task 20 times. The idea is sound. The headline reliability number is a tail count that shows up as two different values in the blog and the paper, and what it mostly measures is the shape of the task set.

Wait, Ask, or Act: What a Proactive Agent Is Allowed to Do Unprompted

15 minute read

· Agents

OpenAI launched dots on 29 September, an agent that works while you are away. Meta’s Muse spent the same fortnight in the news for acting outside what its users thought they had granted. The two products describe similar controls. OpenAI has published its own failure rates, and I have not seen Meta publish comparable ones.

The Harness Tax Shows Up in the Bill, Not on the Scoreboard

20 minute read

· Agents

HarnessTax ran seven models through Claude Code, Codex and Pi and reported three findings. The study’s own chart data, which carries its significance tests, supports the first: the harness moves cost. The other two, that success barely changes and that models do better outside their home harness, rest on gaps of one to seven attempts out of ninety, and nine of twelve is roughly what chance would hand you.

page 1 of 2 · 22 essays on Agents