New agent harnesses scale by experience, yet teamwork and forgetting set limits

Scaling by experience

In a preprint, CUA-Master introduces a nested recursive improvement framework for computer-use agents, using an outer world-data-model flywheel and an inner flywheel coupling open-world seeding with autocatalytic propagation. The former mines authentic workflows as seed tasks from instructional videos and uses visual-semantic dual-path trajectory reconstruction to create reusable task primitives, while a feedback-guided snapshot-rollback retry mechanism improves the efficiency and reliability of trajectory collection in real environments. m:…0792

CUA-Master-27B achieved 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench and 25.0 on Agents' Last Exam, according to the preprint. NorthCUA, in a separate technical report, takes a similar route: an Experience Flywheel supplies executable tasks and curates verified trajectories, recursive self-improvement evolves reusable skills and preserves verified milestones for unsolved tasks, and staged learning plus reinforcement learning build GUI competence, producing a Nano, Mini and Pro model family described as achieving state-of-the-art results among open-source models. m:…0792 m:…4101

CoworkOS, a long-horizon agent harness described in a preprint, separates task strategy from execution management: models handle delegation and acceptance while the runtime maintains state, using a three-layer architecture, hierarchical workspaces and persistent semantic memory so working files, decisions and long-running job logs remain shareable, reusable and resumable after interruption. Human participation enters through conversation and through direct modification of shared artifacts such as CAD models and documents, and automated post-training with GLM-5.3 as controller lifted Qwen3.6-35B-A3B's average score across eight benchmarks from 33.19 to 42.95 over five rounds. m:…3760

The teamwork gap

A second preprint, the Collaborative Intelligence Benchmark, tests the other side of the same ambition: agents as teammates. It contains 240 tasks across eight dimensions of teamwork, drawn from real records and existing benchmarks, from aviation accident reports and clinical cases to coding tasks and negotiations. Every model, across sizes and families, has clear weaknesses as a teammate, the authors write, and general capability does not guarantee good teamwork. m:…4003

The best model scored only 82.2 out of 100, models caved under interpersonal pressure, reaching the best overall deal in only 14% of conflict scenarios, and their pass rate dropped from 97% to 81% when they took over a task in which a weaker model had cheated. Models with similar individual performance scores differed by up to 11 points on collaborative tasks, which means a leaderboard score for solo capability is a poor predictor of how an agent will behave in a team. m:…4003

A further preprint on continual evolution sharpens the warning. Harness self-evolution can improve agent capabilities without modifying foundation-model parameters, and with DeepSeek-V4-Flash-0731, Meta-Harness improved SWE-Bench performance to 94.0% after evolution, but after further evolution on FeatureBench, performance on SWE-Bench dropped to 32.0% by the fifth round, a 66% relative decrease. Evolution on a subsequent dataset can substantially undermine performance on previously seen datasets, the authors write, which is catastrophic forgetting in the harness rather than in the model. m:…4348

Building agents that can be managed

Sierra's redesign of Agent Studio around Ghostwriter, its agent for building agents, points to the operational answer: put the agent-building work in one chat-based session where a team can investigate why resolution rates dropped, examine conversations, update a journey and test the fix, with changes staying in the session until someone reviews and merges them. Release governance adds required approvals and controlled rollouts, so nothing reaches customers until the team says so, and sessions can be shared or forked so colleagues can pick up where another left off. m:…1422

Ghostwriter combines what used to be spread across Explorer, Sierra Assistant and other specialised agents, so teams no longer need to know which tool does what, and it stays with the user as they move through Agent Studio, seeing whether they are looking at a journey, a simulation or a dashboard. Line-by-line comparisons show exactly what changed before anything merges, and a teammate's changes can be sent to Ghostwriter for review. m:…1422

The combined picture is a field learning to manage agents the way it once learned to manage code: scale them through experience, measure them as teammates rather than solo performers, guard against forgetting when they evolve, and keep a human approval gate on what ships. The preprints are early evidence, not settled results, but they agree on where the next failures will come from. m:…0792 m:…4003 m:…4348 m:…1422