In one paragraph
An agent harness is the code around a model: the loop that calls it, the tools it can use, where state is kept and how it runs. These repos differ in where they put the weight. Some are coding-agent CLIs and apps, some are runtimes built on an event log or on cheap idle agents, some tune the harness itself, and a paper formalizes how components compose.
What counts here: frameworks, harness designs and reference stacks for building one agent: its loop, tools, state and runtime.
The main approaches
Harness frameworks
These give you working code for the agent loop, the tools and the state, and differ in what they ask you to build on top.
Coding agents you extend. goose is a Rust agent with a desktop app, a CLI and an embeddable API. It connects to 15+ LLM providers and lists 70+ MCP extensions. pi is a TypeScript monorepo with a unified LLM API, an agent runtime, a terminal UI and a coding agent CLI. It leaves out sub-agents and plan mode on purpose and expects you to add them through extensions, skills and packages shared over npm or git. earendil-works/pi has 122,471 stars, and aaif-goose/goose has 56,710.
Runtimes built around state. infinity keeps an idle agent as plain conversation history, so waiting agents cost nothing until the next event. It also defines a tool protocol that delivers results to agents with no running process. tardigrade folds an immutable event log into agent state through composable atoms. It runs on Cloudflare Workers or a self-hosted Celld fleet.
Platforms and stacks. equin and equin pair an agent manager for parallel sessions with background agents and a shared plugin system, aimed at large apps. singularity is a Bun plugin monorepo with Claude memory and skills directories at the root. ecosystem-overview is a README-only map of a layered stack: Chronicle storage, Membrane for LLM calls, a context manager and the Connectome agent framework.
Harnesses for training and tuning. feather-train is a knowledge-work harness with prod and train profiles, plus packages for rollouts and evaluation. harness2 treats the harness as something to refine for each task while model weights stay fixed.
Composition research
This approach asks how components can be added to and removed from a running system without breaking it. paper is a preprint on spatiotemporal composability. It uses revertible effects and reactive coeffects, and the implementation target is the Cordis meta-framework.
Map of the theme
flowchart LR t["Agent harnesses and frameworks"] t --> f1["Harness frameworks"] t --> f2["Composition research"] f1 --> r1["aaif-goose/goose"] f1 --> r2["earendil-works/pi"] f1 --> r3["anima-research/ecosystem-overview"] f1 --> r4["Conchylicultor/equin"] f1 --> r5["Conchylicultor/singularity"] f1 --> r6["equinai/equin"] f1 --> r7["hydro-project/infinity"] f1 --> r8["clavia-labs/tardigrade"] f1 --> r9["paper-instruments/feather-train"] f1 --> r10["google-research/harness2"] f2 --> r11["cordiverse/paper"]
Where the new ideas are
- infinity: agents that hibernate as pure conversation history, with tool results delivered to agents that have no process.
- tardigrade: agent state derived from an immutable event log through atoms.
- harness2: the harness itself is optimized per task, by proposing edits, probing them and composing a result.
- feather-train: one agent loop with
prodandtrainprofiles, so live use and rollout collection share code. - paper: a formal account of how components can be removed and reconciled cleanly.
- pi: strong defaults with deliberate gaps, filled by shareable packages.
- ecosystem-overview: a branchable record store with undo and redo under the agent framework.
Side by side
| Repo | Approach | Language | Extension mechanism | Where it runs |
|---|---|---|---|---|
| aaif-goose/goose | Harness frameworks | Rust | 70+ MCP extensions | Your local machine, via desktop app, CLI or API |
| earendil-works/pi | Harness frameworks | TypeScript | Extensions, skills, prompt templates, Pi packages over npm or git | Interactive TUI, print or JSON mode, RPC |
| anima-research/ecosystem-overview | Harness frameworks | Not stated | Agent-framework modules for Discord, files, API and MCPL | Not stated |
| Conchylicultor/equin | Harness frameworks | TypeScript | Plugin system of reusable pieces | Not stated |
| Conchylicultor/singularity | Harness frameworks | TypeScript | Bun workspace plugins under plugins/ | Not stated |
| equinai/equin | Harness frameworks | TypeScript | Plugin system of self-contained feature blocks | Not stated |
| hydro-project/infinity | Harness frameworks | Rust | Tools over the Reactive Agent Protocol, MCP compatibility layer | Resident process or serverless such as AWS Lambda |
| clavia-labs/tardigrade | Harness frameworks | TypeScript | Atoms wired into actors with defineActor | Cloudflare Workers or a Celld fleet |
| cordiverse/paper | Composition research | Not stated | Cordis declarative component loader | Not stated |
| paper-instruments/feather-train | Harness frameworks | Python | Paper document skills, prod and train profiles | Local CLI, E2B sandboxes for rollouts |
| google-research/harness2 | Harness frameworks | Python | Eight editable components including skills and plugins | Not stated |
How the idea moved
flowchart LR n1["started Aug 2024<br/>aaif-goose/goose"] n2["started Aug 2025<br/>earendil-works/pi"] n3["started Feb 2026<br/>anima-research/ecosystem-overview"] n4["started Apr 2026<br/>Conchylicultor/equin"] n5["started Apr 2026<br/>Conchylicultor/singularity"] n6["started Apr 2026<br/>equinai/equin"] n7["started Apr 2026<br/>hydro-project/infinity"] n8["started Aug 2026<br/>clavia-labs/tardigrade"] n9["started Aug 2026<br/>cordiverse/paper"] n10["started Sep 2026<br/>paper-instruments/feather-train"] n11["started Oct 2026<br/>google-research/harness2"] n1 --> n2 --> n3 --> n4 --> n5 --> n6 --> n7 --> n8 --> n9 --> n10 --> n11
- started Aug 2024 · goose · Adds: a local agent with a desktop app, a CLI and an embeddable API, 15+ LLM providers and 70+ MCP extensions, plus a guide for building custom distributions
- started Aug 2025 · pi · Adds: a coding agent with deliberate gaps, built from separable packages (LLM API, agent core, TUI) and extended through shareable npm or git packages
- started Feb 2026 · ecosystem-overview · Adds: a layered map of a research stack, covering a branchable event store, a multi-participant LLM layer, a context manager and an agent framework
- started Apr 2026 · equin · Adds: an agent manager and plugin system that let one author’s apps share building blocks, so an improvement to one shows up in the others
- started Apr 2026 · singularity · Adds: a Bun plugin workspace that keeps Claude memory, skills and MCP config alongside app plugins
- started Apr 2026 · equin · Adds: an agent platform for large apps with an agent manager for parallel sessions, background agents that fix and push updates, and reusable plugin blocks
- started Apr 2026 · infinity · Adds: a Rust runtime where waiting agents are only conversation history, with a tool protocol that delivers results to agents that have no running process
- started Aug 2026 · tardigrade · Adds: durable agent state derived from an immutable event log through composable atoms, hosted on Cloudflare Workers or Celld
- started Aug 2026 · paper · Adds: a formal calculus of revertible effects and reactive coeffects, showing that dynamically loaded components can interleave without disturbing one another
- started Sep 2026 · feather-train · Adds: one knowledge-work harness with prod and train profiles, plus packages for sandboxed rollouts and Harbor evaluation
- started Oct 2026 · harness2 · Adds: per-task refinement of an agent’s prompts, skills and tool guidance by proposing, probing and composing edits, with no ground-truth feedback
Easily confused
- An LLM client library. A client library wraps provider APIs. A harness also owns the loop, the tools and the state around those calls. Pi ships a unified LLM API as one package among several, and Goose’s provider support is one part of a larger agent.
- Multi-agent orchestration. A harness is about building one agent well. equin’s agent manager also runs many sessions, but the unit is still the single agent’s loop and runtime.
- A finished coding assistant. Goose and Pi can be used that way. Others, such as Tardigrade and Infinity, are frameworks you build agents with.
- The equin and singularity repos. They share the one-line description “A self-evolving personal OS”. The equin READMEs describe an agent platform for building large apps, with the personal OS as one use of it.
Gaps nobody has filled
- No head-to-head comparison of different harness frameworks on the same tasks. Harness2 compares its refined harness only against base harnesses, and Feather runs its own harness on Harbor benchmarks.
- No shared format for agent state, so an event log, a conversation history and a durable runtime cannot be swapped for one another.
- No measurement of what idle or hibernating agents cost across runtimes. The one figure is Infinity’s README claim of 75k agents in a Raspberry Pi’s memory.
- No link between the composition theory and a harness benchmark, so whether clean removal of components improves agent results is untested.