The terminal is becoming the agent's gym: ReTerm, ClawLoop and Harness^2

Agents building their own gym

The most striking preprint of the four treats environment design itself as a terminal task. In a preprint, the AutoEnvScaling authors describe a pipeline where agents build the reinforcement-learning environments they then train in — a data flywheel with no human in the loop. They report the valid-environment rate rising up to 25 times versus prompt-only generation, cost per valid environment falling up to 75%, and Qwen3.6-35B-A3B gaining 12.2 points on Terminal-Bench 2.1 while serving as both builder and solver. The pipeline generalizes across software engineering, computer use, professional work and GPU kernel programming, per the paper. m:…9949

If that sounds recursive, it is meant to be. The authors state the broader goal plainly: recursive self-improvement. The terminal — with its verifiable outcomes and endless task supply — is the natural place to attempt it, because unlike open-ended chat, a shell either ran the command correctly or it did not. m:…9949

In a preprint, the authors describe what a self-built environment actually looks like: a proposer agent constructs complete Harbor tasks — Docker images, instructions, solutions, tests — with no human assembly. The reported trajectory is steep: Qwen3.6-35B-A3B raised its Terminal-Bench 4.0 score from 1.6 to 3.2 in the loop, and environments generated by a stronger builder lifted a smaller model too — training Qwen3.5-35B-A3B on AutoEnvScaling environments from GPT-5.6-sol improved its Terminal-Bench 2.1 score by 4.3 points. The data flywheel, in other words, has a working prototype. m:…9949

Harnesses that rewrite themselves

Two more preprints attack the harness — the scaffolding around the agent — rather than the model. In a preprint, the ClawLoop authors describe a three-phase loop (task synthesis, training, diagnosis and feedback) that generates its own training data and refines performance with no external teacher model, reporting gains on SkillsBench and ClawGym-Bench and successful transfer of refined tasks from a 27B loop model down to 4B and 9B models. The economic argument is explicit: teacher models are the expensive dependency, so cut them out. m:…0658

In a preprint, the Harness^2 authors go one level meta: their framework recursively refines the execution harness itself, at both task level (via a contrastive Propose-Probe-Compose step) and harnessing level, without ground-truth verifiers — which matters because open-world tasks lack them by definition. They report up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB for GLM-5 and Gemini 3.5 Flash. The throughline: the harness is becoming a trainable artifact, not a hand-written config. m:…5030

The transfer results are the part to take seriously. In a preprint, the ClawLoop authors report that tasks refined by the 27B loop model, transferred to Qwen3.5-4B and 9B, produced gains across all benchmarks — the data quality survives the downscale, which is what makes teacher-free training economically interesting rather than merely elegant. Their synthesis step also grows difficulty conditioned on diagnosis, and validates tasks both with and without an agent in the loop. And in a preprint, the Harness^2 authors report that gains increase as the task-level recursion budget increases — the loop compounds, at least within the tested horizon. m:…0658 m:…5030

Interfaces built for agents, not humans

The fourth preprint asks an embarrassing question: why do we make agents use interfaces designed for people? In a preprint, the authors introduce DeskIF-Bench — 255 artifact-verified tasks across six desktop applications — and show that interface properties like discoverability, granularity and observability drive agent success. Their ANIS framework then synthesizes agent-native interfaces from software evolution history, reporting a 3.8 percentage-point success gain for DeepSeek V4 Flash and roughly 73% lower online token usage than native SDKs. Their slogan is the takeaway: do not just teach agents to use existing tools — build tools agents can use effectively. m:…5915

For developers, the actionable pattern across all four is to stop treating the environment, harness and interface as fixed. Each preprint moves one of them into the optimization loop, and each reports that the gains compound. If you run terminal agents in production, the cheapest experiment on this list is probably ANIS-style interface synthesis — a 73% token cut pays for itself fast — while AutoEnvScaling's self-built environments are the one to watch for training-data strategy. m:…5915 m:…9949

One more number from the ANIS paper, in a preprint: goal-level commands — letting the agent express intent at the task level rather than clicking through steps — raised success to 74.5% for DeepSeek V4 Flash and 67.4% for Claude Haiku 4.5. The authors also report that newer software releases consistently outperform older ones across both models and six applications, a quiet argument for keeping your desktop fleet current if agents are going to drive it. Interface quality turns out to be a moving target — and it moves in the agents' favor. m:…5915