DeepSeek’s Harness Release Turns AI Agent Tooling Into a Three-Way Bet

On August 13, 2026, DeepSeek published a GitHub repository called deepseek-harness with a one-line pitch: “Everything is a Plugin.” According to freeCodeCamp, the repository passed 95,386 stars and 8,826 forks within two days, one of the fastest growth curves any developer tool has posted this year. A month later, sxz.io independently confirmed via GitHub’s own API that the repository has grown to more than 222,000 stars and 26,000 forks, and its release has settled an argument that had been simmering in AI developer circles for most of 2026: the word engineers use for the software wrapped around a language model has quietly changed from “framework” to “harness,” and at least three incompatible answers now compete for what that harness should actually do.

What an Agent Harness Actually Does

A language model, by itself, does one thing: given a stream of text and a list of available tools, it predicts what to say or which tool to call next. Everything else, the loop that calls the model repeatedly, the code that executes each tool call, the memory that decides what survives into the next turn, and the sandbox that limits what a tool call can touch, is the harness’s job.

A typical exchange runs in a fixed cycle. The harness assembles a prompt from the user’s request, the system instructions, and the tool schemas the model is allowed to call. The model responds with reasoning text plus, usually, one or more tool calls. The harness executes each call, ideally inside a sandbox, and captures the result. The harness appends that result back into the conversation and calls the model again. The loop can repeat for dozens of turns before the model produces a final answer, the harness hits a turn or cost limit, or a human interrupts it.

That loop, often called the ReAct loop after a 2022 research paper by Shunyu Yao and coauthors that first described interleaving reasoning traces with actions as a single process, is the one piece every harness in this piece shares. What differs, and what determines whether a given harness is actually good to build on, is everything wrapped around the execute-and-append step: how much of the wiring is a swappable plugin, how permissive the sandbox is, and how much state survives from one session to the next.

Why “Framework” Stopped Being the Word

From 2023 through 2025, “framework” was the dominant term for this category. Tools like LangChain, AutoGen, and CrewAI were frameworks in the traditional software sense: a team imported components, wrote its own orchestration logic, and assembled the agent loop by hand. A framework handed a team building blocks; the team still had to build.

The clearest evidence the vocabulary has actually shifted, not just among outside commentators but inside the companies that make these tools, sits in a single line of text. LangChain, one of the companies most responsible for making “framework” the default term in 2023, now describes its own newer agent product, deepagents, on GitHub as “The batteries-included agent harness.” sxz.io independently confirmed that description is the repository’s live title as of this writing. A company that helped build the original framework era now brands its own product with the word that replaced it.

Three Incompatible Bets on How a Harness Should Work

Beneath the shared vocabulary, the agent tools that have gained the most traction in 2026 disagree sharply on how a harness should actually be built. sxz.io independently pulled current statistics for five widely discussed projects directly from GitHub’s REST API rather than relying on secondhand star counts, and the resulting picture cuts against the assumption that whichever tool gets the most press coverage is winning.

Bet One: Everything Is a Plugin

DeepSeek’s own description for deepseek-harness, published under an MIT license on August 13, 2026, is blunt: “Everything is a Plugin.” Models, tools, sandboxes, and the user interface are all swappable components rather than fixed choices baked into the core. As of this writing, GitHub’s API puts the repository at 222,015 stars and 26,344 forks, a curve that is still climbing a month after launch and remains the fastest-growing of the five projects in this comparison by a wide margin.

Bet Two: A Small, Fixed Set of Mechanisms, Executed Well

Pi, published by the small studio Earendil Works and closely associated with engineer Mario Zechner, takes close to the opposite position. Its GitHub description calls it an “AI agent toolkit: unified LLM API, agent loop, TUI, coding agent CLI,” built around a small number of built-in tools rather than a plugin marketplace, with anything else added as an opt-in TypeScript extension. GitHub’s API shows the repository was created in August 2025, roughly a full year before DeepSeek’s launch, and has grown to 104,516 stars and 13,117 forks through what looks like steady, compounding adoption rather than a single viral spike, a genuinely different growth shape from deepseek-harness’s two-day surge even though both projects now sit in the same six-figure range.

Bet Three: Memory That Compounds

The largest of the five repositories by star count is not the one that generated the most 2026 press coverage. Nous Research’s hermes-agent, described on GitHub simply as “the agent that grows with you,” has reached 245,004 stars and 50,886 forks, according to GitHub’s own API as of this writing, ahead of even DeepSeek’s viral release. Where the other two bets treat each agent run as a disposable instance, Hermes Agent is built around one persistent identity that, per its own README, is reachable from Telegram, Discord, Slack, WhatsApp, Signal, and email from a single gateway process, and that is designed to accumulate skills and context over time rather than reset between conversations. The project’s own GitHub organization backs that framing with a working ecosystem rather than a slogan: a hermes-agent-self-evolution repository that uses DSPy and GEPA to let the agent optimize its own prompts and skills, a hermes-plugin-snyk repository that wires in Snyk’s code and dependency scanning, and a hermes-paperclip-adapter that connects Hermes to Paperclip, a separate app for managing teams of AI agents at work, letting it run there as, in the adapter’s own description, “a managed employee.”

Where Claude Code and LangChain Fit

Two more data points round out the picture. Anthropic’s own Claude Code, the terminal-based coding agent tied directly to Anthropic’s models, was created in February 2025 and has grown to 144,891 stars and 23,136 forks, positioning it between Pi’s steady climb and the two viral outliers. It represents a fourth category worth naming on its own: a harness built and maintained by the same company that ships the underlying model, rather than by an independent team plugging into whichever model a user chooses.

LangChain’s deepagents, meanwhile, is smaller than any of the other four by star count (29,356 stars and 4,112 forks as of this writing) but arguably the most symbolically important. It is direct, primary-source evidence that the framework-to-harness shift described above is now baked into official product naming, from the same company that helped make “framework” the term everyone used in the first place.

The Shared Plumbing Underneath: MCP and the Agent Loop

None of these three bets exists in total isolation. Most of the harnesses named here, including Claude Code and deepagents, along with a fast-growing share of the wider ecosystem, can speak the Model Context Protocol, the open standard Anthropic introduced in late 2024 for connecting a model to external tools and data sources in a vendor-neutral way. A tool built to that specification for one harness will generally work with another, which means the “everything is a plugin” bet and the “small, fixed set of mechanisms” bet are not as incompatible in practice as their design philosophies suggest. A Pi user can often reach the same external tool a DeepSeek Harness plugin exposes, just through a narrower, more opinionated interface.

Underneath the protocol layer, every one of these projects still runs the same request, tool call, response loop that traces back to 2022. The disagreement that actually matters for a team choosing between them is not about that loop. It is about how much of the wiring around it a team wants to own.

Why the Terminology Shift Matters for Teams Choosing Tools

For a team evaluating agent tooling in September 2026, the practical question was never framework versus harness. It is which of these three bets matches the team’s own appetite for maintaining infrastructure. A fully pluggable harness like DeepSeek’s offers the most flexibility but also the largest surface area to secure and keep patched, the same permission-boundary tradeoff sxz.io has covered directly when reporting on bypassable agent-autonomy flags such as Claude Code’s own dangerously-skip-permissions mode and on scoped authorization tokens built specifically to limit what a compromised agent tool call can reach. A small, fixed toolset like Pi’s is easier to audit precisely because there is less of it, at the cost of writing a custom extension for anything the built-in tools do not cover. A persistent, self-evolving identity like Hermes Agent’s promises more compounding capability over time, but it is also a harder target to fully audit at any single point, since the thing being reviewed today may have rewritten its own prompts or skills by the time anyone checks again.

The word “harness” won because it describes what all three of these projects are actually doing more honestly than “framework” ever did: not handing a team a box of parts, but wrapping a specific, opinionated shell around a model and betting that the shell is the right one.