Sakana’s MASS reported 1.2 to 1.6 times more performance per output token on four open-ended benchmarks after two revision cycles. On the coding tests, it scored 0.96 times its base model on Terminal Bench 2.0 and 0.94 times on SweetBench Verified.
Revising the workflow helped in the open-ended tests. It did not help on those coding tests.
The Workflow Changed, but the Model Stayed Fixed
A harness is the program around a model. It manages what the model sees, which tools it uses, and how work passes between steps. Changing that workflow can change how an agent tackles a task even when the task-solving model stays the same.
One paper kept the task-solving model fixed and trained a smaller model to revise its harness. Across 21 puzzle families it had not trained on, the revised harnesses averaged 0.62; a larger teacher averaged 0.56. That was a puzzle-solving result after one revision. It does not establish how the approach would perform as a coding agent.
MASS tried a related approach with teams of agents. A team does the task, a model judges the result, and the workflow changes if the new version wins. In MASS, the same model does the work and judges it. The coding scores matter because they show that a workflow which helped on the reported open-ended tests did not carry that gain into both coding tests.
A revision needs a way to earn its place. Anthropic’s evaluation guide describes code checks, model judges, and human spot checks. It also advises grading what an agent produced rather than requiring one exact sequence of steps. That keeps the evaluation focused on the result, even when the route to it changes.
Reviewing the Work Takes Its Own Workflow
For one commentator, the bottleneck has shifted from writing code to understanding what an agent returned. They use diagrams and interactive HTML to inspect changes and make decisions. After some AI-made physics demos were wrong or confusing, they added separate clarity and physics reviews.
Another example put a GitHub sign-in change, a mockup, checks, and tested user flows on one review page. The system asked for approval because the change involved authentication. An Anthropic engineer described using Claude for stand-up suggestions, then asking it to add numbers and a chart to a launch review they had written. In each case, a person still had to decide what to do with the output.
The work also has to survive a handoff. Beacon says its open-source project turns sessions from Claude Code, Cursor, Codex, and more than 20 other harnesses into reusable skills. A separate tip recommends a HANDOFF.md with objectives, decisions, changed files, and next steps when switching agents. Beacon’s claim describes what its tool is meant to do; its reliability at preserving that context was not tested here.
What an Agent Leaves Running Locally
Cloud model work can still leave a browser, development server, or build tools running on your computer. One commentator suggests checking memory pressure and inspecting what an agent left open before buying more hardware or deleting files. For a dozen document jobs, their example starts with one or two heavy steps at a time.
Their suggested Mac starting point is 24 to 32 gigs of memory unless usage is especially heavy. They say 16 gigs can work with more active management. The practical question for that setup is what runs locally, when it runs, and whether finished work has stopped.
Storage reports show why it is worth checking. One public Codex report described a crash folder over 50 gigs. Another described more than 100 gigs of Git objects from turn checkpoints. A separate reporter measured roughly 37 terabytes of SSD writes over 21 days during an earlier Codex logging issue. Those figures came from particular setups; Codex later shipped logging fixes and added reclamation of unused log database space.
Permissions Can Remain After the Job
A tool built with Claude Code uses Android Debug Bridge to disable unwanted Android TV packages for a user profile. It backs up the package list, asks which streaming apps to keep, checks steps with screenshots, and provides a restore script. It supports Android TVs, but not Samsung, LG, or Fire TV.
Setup requires the TV to approve a computer for debugging, and the tool’s author advises turning debugging off afterward. That approval limits a comparison with Anthropic’s worm test, which used eight emulated Android machines with debug ports open without a login. Only GLM 5.2 reached all eight. Anthropic calls its strongest result an upper bound on real-world success; the test does not show the same thing would happen on a TV that asks for approval.
A separate research report says poisoned repositories can steer Claude Code and Codex CLI into granting broad, lasting permissions. The concern is access that remains after the immediate job. And for workflow revision itself, the unresolved test is still the coding one: will it improve the tests where MASS scored below its base model?