Why OpenAI, Claude, Google, and DeepSeek All Bet on the Same Harness Layer

01 OpenAI makes the Harness a reusable runtime

OpenAI describes the Codex Harness as the core Agent Loop that coordinates user input, the model, and tools across Web, CLI, IDE extensions, and desktop apps. The App Server, announced in February 2026, exposes a stable JSON‑RPC interface and manages long‑lived threads, turns, and items, preserving state across page closures or network loss.

02 Anthropic, Google, and DeepSeek productize the layer outside the model

Anthropic : Claude Code’s Harness is exposed via the Claude Agent SDK, allowing sub‑agents, background tasks, and plugins to run in any development environment. Dynamic Workflows (May 2026) let Claude plan large tasks, run many sub‑agents in parallel, and aggregate results.

Google : At I/O 2026, Antigravity 2.0, CLI, SDK, and Managed Agents share the same Harness. The CLI uses the same Harness as Antigravity 2.0, the SDK offers programmatic access, and Managed Agents expose the Harness as a Gemini API, enabling custom deployment and co‑optimization with Gemini models.

DeepSeek : The Developer Preview presents a plugin‑centric design—"Everything is a Plugin"—where the core loads and composes plugins for model, tools, skills, session, sandbox, storage, loop, scheduling, and UI. The model itself is a plugin, allowing flexible recombination.

03 Why Coding Agents were the first to close the Harness loop

Coding environments provide a natural verification loop: code lives in a file system, shells execute commands, compilers and tests return immediate results, and Git tracks changes. This cheap, fast feedback lets agents iterate through observe‑modify‑execute‑verify cycles, making the Harness viable early on.

When extending to non‑coding tasks (customer service, research, ops), verification becomes harder because there is no compiler or unit test. The next generation of agents will need stronger validators, approval mechanisms, audit trails, and rollback capabilities.

04 Harness ≠ Workflow; heavier is not always better

Workflow describes task sequencing; an Agent Framework/SDK helps developers build agents; Harness is the runtime that continuously manages context, state, tool execution, sandboxing, permissions, recovery, validation, and observability. Over‑engineering the Harness can hinder performance, especially as models improve and require fewer constraints.

Anthropic’s experiments show that mechanisms like Context Reset become obsolete after model upgrades, and Sprint‑style task splitting can be simplified when models handle long‑term planning better.

05 The next competition: Model‑Harness joint optimization

Agent Lightning v1.0 (released 18 August) introduces "Harnessed Agentic RL", where the deployment‑side Harness participates in the training loop. Using ~6 000 samples, Qwen‑3.5‑9B’s SWE‑bench Verified score rose from 41.8 % to 56.4 %.

Subsequent work on Harness Continual Learning shows that while the base model can stay frozen, the Harness can keep learning via updated task interfaces, memory, capability maps, and routing rules. However, changes to the Harness can also cause "harness‑level forgetting", indicating that agent competence now resides partly outside model parameters.

Overall, the competitive edge shifts from owning the default Model API to providing a stable, observable Agent Runtime that integrates sandboxing, skills, sub‑agents, permissions, and validation as default infrastructure.