Skip to main content
Search

# Agent Harness

29 related articles

Subscribe to updates
  1. arXiv

    Command-line agents often report skipped work as success: Receipt reads kernel wait states and cuts false successes from 34 to 9

    When LLM agents operate Web services from the command line, they assume every tool call returns within a bounded time with the requested action finished; but many commands ask a question first, and a harness that closes input can let a program return status 0 without doing the work, while an open input makes it wait until a time limit. Receipt keeps input open, reads from system calls what each process waits for, and returns that state to an unmodified agent; across 5 LLMs it lowers false successes from 34 under closed input to 9 and uses 24.4% fewer tokens than a plain terminal.

  2. arXiv

    An agent writes its execution experience into reviewable skills: DrugHarness designs and synthesizes an HPK1 compound with 14.19 nM cellular IC50

    Computational drug-discovery agents usually rebuild their tool calls for every task, so a successful strategy rarely carries over. DrugHarness places more than 80 scientific operations and biomedical resources in one workspace, induces action hypotheses with explicit applicability conditions from execution trajectories, and admits them as skills only after an independent review agent checks sources and runs targeted trials when needed. On MolBench, the configuration without evolution reaches 91.9% affinity-comparison accuracy, 10.8 percentage points above reported MolClaw-CC results; in retrospective evaluation, frozen learned skills raise Qwen's optimization success from 79.5% to 87.2%.

  3. arXiv

    Mobile robots skip the prebuilt semantic map: Feijoa grounds instructions at query time and hits 94.2% exact match on 103 queries

    When a mobile manipulator is told to put the toys away in the toy box, it usually relies on a task-agnostic semantic map that may lack the attributes, affordances or concepts the instruction needs. Feijoa has a language-model agent query a posed RGB-D scan after the instruction arrives, inspect the returned candidates and rewrite its queries until the reference resolves. On 103 open-ended grounding queries it reaches 94.2% exact match, against 68.0% for the best prebuilt representation, TextMap; on 50 real Spot trials it completes 35 tasks, 17 of them after recovering from a failed tool call or verification check.

  4. arXiv

    Robot arms miss grasps by millimeters; Auriga recovers 68.3% of the planning-execution gap

    Vision-language models can read a task yet still fumble the grasp: on RoboDojo, removing arm execution raises the same model's success rate from 22.5% to 76.2%, locating the bottleneck in physical execution. Auriga separates model decisions from robot execution, with the model keeping an explicit task belief and every motion passing a single admission gateway that previews the trajectory, checks stage tolerances during execution, and stops on divergence. Frozen for evaluation, Auriga reaches 57.7% success and 64.0 Score across 42 tasks, ranking first, and 73.2% on RoboCasa365.

  5. T3 Code puts your local coding agents behind one console, reachable from phone, web and desktop

    T3 Code released an agent harness control surface that drives coding agents already installed on a user's machine from one iOS, Android, web and Electron desktop interface. It does not train a model or replace those agents: users must first install and authenticate at least one supported provider, such as Codex, Claude, Cursor, Grok Build, OpenCode, Antigravity or fx, and T3 Code then controls it. The project describes itself as very early, warns of bugs, and is mostly not accepting contributions.

  6. pingdotgg/t3code

    T3 Code puts agent control on your phone: six coding agents already logged in locally can be driven remotely

    T3 Code has released a control surface for coding agents, with iOS, Android, web and Electron desktop clients, that drives agents already installed and authenticated on the user's own machine. It supports six providers, Claude Code, Codex, Cursor, Grok Build, OpenCode and Google Antigravity, each of which must first be set up locally. The project says existing options such as the Codex desktop app, Conductor, Claude Desktop and Cursor Glass did not meet its bar, and describes itself as very early with bugs to expect.

  7. arXiv

    Image-generation agents stop merely retrying: failure experience passes a counterfactual check before entering the skill library, reaching 86.1 on GenEval2

    Agentic image-generation pipelines have largely been read-only generate-verify-rewrite loops that discard repair experience or admit ungated updates into a skill library. VEvo-Harness expands each failed verdict into a hierarchical verification tree to locate root causes, then admits only causally valid skill patches through counterfactual dual-arm image replay, with memory organized into working, experience and pattern tiers. On the Qwen-Image-2512 backend it reaches 86.1 on GenEval2, 14.5 points above GenRouter, while continuing to evolve.

  8. Cloudflare

    Your agent's browser can now open private sites: preview.internal keeps a sandbox dev server off the public internet

    Cloudflare added internal hostnames to the persistent browser in its Agents SDK: give the browser a name such as preview.internal and requests to it go to the agent itself rather than the internet. The agent can then check a dev server in its sandbox, a folder in its workspace, or another Worker in the same browser it uses for the web, with nothing exposed publicly. Each internal host has a fetch function running in the Durable Object, so it can use that object's bindings and state, and a host can carry a target schema so the model chooses which port or folder it serves.

  9. CAR signaling domain cut to 45 amino acids: Canelé's agent-designed receptor beats BBζ on surface expression

    Canelé used an AI-agent-orchestrated rational design loop to refactor the CAR intracellular signaling domain, compressing the second-generation BBζ domain into a 54-amino-acid design called Stoat and then a 45-amino-acid variant, Happy Stoat. At the same mRNA dose, Happy Stoat puts more CAR on the T-cell surface than both BBζ and Stoat and releases more cytokine. Across five donors, Happy Stoat matched BBζ's B-cell depletion at lower mRNA doses.

  10. More agents isn't always better: folding specialist subagents into one loop often raised quality and cut latency

    In enterprise agent deployments, folding routers, planners and domain subagents into a single main loop often beat both a one-prompt-for-everything design and a subagent orchestration design on quality, frequently at lower cost and latency per task. Subagents still fit when context bloat degrades a single loop, when tasks need parallel fresh context windows, or when tool choice and instructions conflict. A guided workflow lets teams try the migration on a mock system first.