google-research/harness2

This page introduces an external open-source repository. It is not an HDATF product.

What it does

This research framework helps agent developers refine an execution harness without changing model weights. Given a task and a base harness, it proposes edits, probes them on that task and composes refinements from execution contrasts. Editable components include prompts, memory, tool guidance and skills. Parallel recursion explores candidates while sequential recursion carries refinements forward. Adopted harnesses and evidence-supported lessons update the improvement harness for later tasks. Outputs include deliverables, adopted configurations, reflections and evaluation records. It targets Linux and Python; model experiments additionally require credentials, an executor and benchmark-specific dependencies. The local demo needs no model credentials, but model experiments incur API costs. Installation does not automatically improve an existing agent. Executor network access and Codex's disabled sandbox require separate controls, and this is not an officially supported Google product.

License

Apache-2.0 Permissive, with a patent grant. Commercial use is allowed; keep the notices and state your changes.

More in this category

• anthropics/claude-quickstarts
  Claude Quickstarts is an official collection of example projects for developers building with the Claude API. It is not one finished product: developers select and extend examples for support, financial analysis or computer and browser interaction. Depending on the project, inputs are questions, documents or task instructions, and outputs include answers, tool results and application interfaces. The autonomous-coding example describes separate initializer and coding agents with progress retained across sessions. Managed Agents examples additionally cover messaging integrations, daily briefs and knowledge wikis. The inspected minimal browser example drives one Chromium tab over CDP and connects tool calls to the model runner. Model execution needs Claude API authentication, example-specific Python or Node.js environments and external-service configuration. Installing the repository neither activates every example nor connects business systems. The minimal browser driver is not production code and omits capabilities such as typing and uploads. Page content can influence the agent, requiring disposable browser profiles, isolation and network controls.
• SakanaAI/AI-Scientist-v3
  The Agentic Scientist, in the repository AI-Scientist-v3, is Sakana AI's Apache-2.0 Python pipeline in which coding agents carry a task through experiments to a paper. It is for computational research with a program to improve and an evaluator that scores it, and the examples are circle packing on a CPU and GRPO training on two GPUs. Input is a task directory with task.md, an initial program and an evaluator, and drafting a task from a short brief alone is experimental, works for small tasks such as circle packing and may fail on larger ones. Six phases run in order: task drafting, investigation that repeats ideas and experiments, ablations against baselines, a sealed final evaluation, post-hoc analysis and paper writing. Each run creates a directory with the task snapshot, attempts and results, cost records and a LaTeX paper with its PDF, and a run inspector shows it in the browser. The paper phase links each claim to evidence files with hashes and gathers references from Semantic Scholar, OpenAlex or Google Scholar, and it fails when citations or evidence links do not pass validation. It drives Claude Code, Codex, Cursor, OpenCode, Gemini and Pi through the Headless CLI, and an optional review phase sends the finished paper to an OpenAI-based reviewer model. Getting started needs Python 3.11 or later with uv, Docker, and API keys or a Claude login in .env, GPUs are needed only for GPU presets, and agsci doctor is the first check. The README warns that agents run any command with full tool permissions, inherit the API keys and are steered by the files they read, and it asks for runs inside Docker. One GRPO evaluation trains for one to two hours, and the cost files separate provider-reported costs from list-price estimates, leave hardware costs out and are run telemetry, with provider invoices as the billing record.
• perplexityai/SPACE
  SPACE is Perplexity's Apache-2.0 platform, written in Rust, that runs agent code isolated in Firecracker microVMs (lightweight virtual machines). It is for applications that need a separate execution environment for an agent, a user workspace or a code-running job, and the application decides what to run and what to keep. A template is the starting environment, and a sandbox is an instance of it that runs several commands in its own VM and writable filesystem. Applications create sandboxes and command sessions through an HTTP API and poll an operation ID until each sandbox lifecycle change completes, and command output and the exit result arrive as server-sent events over a separate connection with a session-scoped token. Pausing keeps memory on the worker and suspending saves a snapshot, and a sandbox suspends by default after six hours without activity. An E2B compatibility wrapper lets existing E2B clients use SPACE as their backend, and its tested baseline is the Python packages e2b 2.21.1 and e2b-code-interpreter 2.6.2. Network policies allow or deny the destinations a sandbox can reach, and an optional credential rule, off by default, has the TLS-intercepting gateway add a credential header to requests for a named host. The local setup builds a Kind (Kubernetes in Docker) cluster on an apt-based Linux x86_64 host with KVM, and macOS is outside the supported setup. The repository description speaks of pluggable runtimes, while the documented implementation currently covers Firecracker on KVM only. Version 0.8.0 is the public baseline, breaking changes remain possible until 1.0, and only the latest minor release receives security fixes.
• wangmiaozero/pi-harness
  Pi-Harness is an AGPL-3.0-only Electron desktop app that wraps Pi Coding Agent in a visual workspace while the native Pi runtime still runs the agent. It runs on macOS, Windows and Linux, and sessions, configuration, tools, Skills and extensions stay compatible with Pi, so work can continue in Pi outside the app. The workspace covers conversations, projects, a file browser and editor, Git diffs, history, commits and worktrees, model providers, and Skills, packages and MCP. Its Harness view records trace, replay, a run tree of forks and retries, tokens, cost and duration, and adds tool policies, per-run budgets and checkpoints. In the source, evaluation is computed from run evidence such as test, lint, typecheck and build commands with their exit codes and the Git state, with no AI judge. The default policy denies file changes outside the workspace and force pushes, asks before deletes, pushes and network commands, confirms commands such as rm -rf, and leaves budget limits empty. Agent teams assign tasks with dependencies and priorities, run them in order or in parallel, pass results on, use review gates and keep changes apart in Git worktrees. A history view reads the local conversation records of Claude Code, Codex, Cursor, Cursor Agent, Gemini CLI, OpenCode and Pi read-only and continues a task in a new Pi chat through a handoff. Installers are on GitHub Releases, the latest being 2.0.0 of 14 October, which follows Pi 1.2.0's breaking changes, and macOS community builds may be unsigned. On Windows and Linux, API keys are encrypted in the app's own vault, while the value written into Pi's models.json is the plain key.
• onehumanbeing/esp32-iphone-agent
  ESP32 iPhone Agent is an MIT-licensed project made of a Python server and Arduino firmware that lets an AI vision model operate a real iPhone without a Mac, Xcode, a developer account, an app on the phone or a jailbreak. It is for developers and hobbyists who want to automate an iPhone through the same paths a person uses, the screen and a pointer, and the README asks users to automate only devices and accounts they own or are allowed to use and to follow each app's terms. The phone mirrors its screen over AirPlay to UxPlay on a macOS or Linux computer, where GStreamer turns it into JPEG frames, and an ESP32-S3 board plugged into the phone's USB port appears as a standard USB mouse and keyboard. With iOS AssistiveTouch (an accessibility setting that shows an on-screen pointer) turned on, a click becomes a tap and a drag a swipe, and the board carries out only simple one-line commands such as click and type received over 2.4 GHz Wi-Fi. All decisions are made by a FastAPI server on the computer, and in that code the web API and the board port listen on every network interface by default without a login, with commands sent as plain text over TCP. A user types a goal into the web page on port 8890, a planner model writes a plan of up to 6 steps, and an actor model returns up to 6 actions such as tap and swipe in coordinates from 0 to 1000 for each new screenshot. Because the pointer moves relatively, the server calibrates it by wiggling it and finding it on the mirrored screen, checks after each action whether the screen changed and tells the model, and re-plans when the same action repeats without effect. A run ends at 40 steps or after 3 errors in a row, and the web page shows a step wall with the screenshot behind each decision and its action drawn on top, plus an emergency stop that the Esc key also triggers. Setup needs an ESP32-S3 board flashed with the firmware (the tested board is the Seeed XIAO ESP32S3 at about 7 dollars), a USB-C iPhone or a Lightning iPhone with a camera adapter, and Python 3.10 or later. The model can be any OpenAI-compatible chat API that accepts images, with separate models possible for the planner and the actor, and since every call carries the current screenshot, with a cloud API whatever the phone screen shows goes to that provider. Development and timing used a local model on two NVIDIA DGX Sparks (about 1.3 seconds per action and 2 seconds per plan), and the bundled serve folder holds an experimental server whose full agent run is still untested. The README lists its limits: iOS keeps passcode entry, Apple Pay and DRM video out of mirroring so the agent cannot see them, mirroring stops when the phone locks, typing is ASCII only, and landscape apps remain unsupported.

Only repositories in the ranked Trendshift lists are included, and the lists are used only to find candidates. We do not copy their ranks. Descriptions, licenses and star counts come from each GitHub repository. The notes are our own reading. We have not tested these projects, and a place on a trending list does not prove quality.