HarnessRouter is an open-source project by Richard Song (songrenchu) that provides a unified API for integrating different AI coding agent harnesses—such as OpenAI's Codex and Anthropic's Claude Code—as interchangeable backends. The project aims to simplify building applications that can route tasks across multiple agent runtimes without needing to build bespoke integrations for each. The web snippets confirm that the "harness engineering" concept is gaining industry traction in 2026, with major players like OpenAI and Martin Fowler publishing on the topic, and meta-harness tools like Omnigent also emerging, though HarnessRouter specifically appears to be an early mover in standardizing this integration layer. The snippets don't show independent adoption numbers or direct endorsements from the major harness providers, only the project's existence and the broader trend.
HarnessRouter directly implements Scott's agent-native computing architecture: it provides the machine-native abstraction layer between agent runtimes and consuming applications, exactly as the Agent-Native Computing framework prescribes — 'machine-native in the middle, human-legible at the boundaries.' This is not merely an example of the pattern; it is a practical instantiation of the architectural separation Scott has argued for. The project also converges with his work on multi-backend routing (his own `ask` terminal agent has a multi-model compatibility layer), his agent-loop architecture, and his code-first architecture position ('replace model-facing tool catalogues with a sandboxed execution environment'). The radar already tracks this story under 'agent-harnesses' (125 episodes) and the prior Omnigent case — this is a new development within an already-open case, not a fresh discovery.
ip:framework.agent-native-computingip:framework.code-first-architectureip:framework.agent-loopdev:project.askradar:concept.agent-harnessesradar:concept.agent-harnessradar:karpathy-claude-md-rulesradar:agent-handoff-protocol-adoptionradar:dropstone-persistent-agent-runtimeradar:agent6-jailed-state-machine-harness
queries asked of Scott's wikis
- Scott's position on agent-harness abstraction layers vs bespoke runtimes
- Scott's work integrating multiple coding agent backends into a single system
- Scott's views on the harness engineering paradigm shift in 2025-2026
- Scott's projects that route tasks between Codex, Claude Code, or other agents
- Scott's framework for evaluating when to adopt a unified API vs maintain separate integrations
- Scott's prior commentary on Richard Song or HarnessRouter specifically
2026-08-29T00:23:59Z
No HarnessRouter adopter, provider integration, portability result, or benchmark followed the independent bespoke implementations. The broader cross-harness pattern is real, but it is currently fragmenting around point solutions rather than validating HarnessRouter as the canonical API.
2026-08-26T23:25:27Z
Agent-hop adds another independent implementation showing that live cross-harness context transfer is technically practical, strengthening the broader interoperability pattern. Its bespoke approach also reinforces that implementations are fragmenting without adopting HarnessRouter, leaving the canonical-API and adoption hypothesis unvalidated.
2026-08-26T23:23:17Z
evidence attached: hn.story.49457119 — Agent-hop is a concrete first-party artifact for switching between Claude Code, Codex, and other agent harnesses while preserving live context, directly testing harness interoperability.
2026-08-26T15:35:47Z
The refreshed discussion adds no distinguishable evidence beyond the already-priced bespoke Claude–Codex workflow. Independent convergence supports the broader orchestration pattern, but HarnessRouter still lacks an adopter, provider integration, portability result, or benchmark validating its canonical API.
2026-08-26T12:34:10Z
The refreshed user comment adds concrete workflow detail—Claude planning, Codex review, and handoff criteria—making cross-harness orchestration slightly more credible in practice. It remains another anecdote about a bespoke setup, not evidence of HarnessRouter adoption, portability, or value as the canonical abstraction.
2026-08-26T04:37:05Z
The user report adds a second independent indication that Claude Code–Codex orchestration can work in practice, strengthening the broader harness-as-backend pattern beyond a lone implementation artifact. It still provides no HarnessRouter adoption, canonical-API validation, portability proof, or provider participation, so the core hypothesis remains untested.
2026-08-26T04:23:22Z
evidence attached: reddit.post.1vyloyk — Claude Code dynamically orchestrating Codex is independent evidence that one coding-agent harness can practically use another as a backend.
2026-08-25T14:40:03Z
The 48-hour quiet leaves the independent convergence signal isolated: HarnessRouter still has no user adoption, provider integration, portability proof, or benchmark validating its canonical API. Cool the case until a concrete implementation or adopter appears.
2026-08-23T13:38:02Z
A separate Codex–Claude Code orchestrator is the first concrete sign of independent convergence on cross-harness composition, modestly strengthening the broader harness-as-backend premise. It does not use or validate HarnessRouter’s canonical API, and the thin artifact provides no portability, usage, or adoption evidence.
2026-08-23T13:22:58Z
evidence attached: hn.story.49408449 — Direct implementation evidence for embedding Codex and Claude Code as interoperable harness backends, materially bearing on adoption of harness-as-backend workflows.
2026-08-23T12:33:06Z
No independent implementation, user, benchmark, or provider integration has appeared since the architectural premise strengthened. HarnessRouter remains a relevant but unvalidated cross-harness abstraction, and the case should cool until concrete adoption or portability evidence emerges.
2026-08-21T12:25:20Z
The newly attached item is duplicate coverage of OpenAI’s already-priced Codex platform announcement, while refreshed comments sharpen portability and lowest-common-denominator concerns without adding independent implementation or adoption. Harness-as-backend is increasingly credible as an architecture, but HarnessRouter’s canonical cross-harness layer remains unvalidated.
2026-08-21T12:22:54Z
evidence attached: hn.story.49386690 — shared external link with case evidence
2026-08-21T04:34:43Z
OpenAI’s first-party positioning of Codex as an open harness developers can build on materially strengthens the architectural premise that established agent harnesses can serve as product backends. It does not yet validate HarnessRouter’s canonical cross-harness layer, portability, or independent adoption.
2026-08-21T04:22:44Z
evidence attached: hn.story.49383362 — OpenAI’s first-party Codex platform announcement materially supports the case for treating established agent harnesses as reusable backends.
2026-08-20T11:32:42Z
The attached item suggests adjacent interest in decoupling coding agents from model backends, but provides no details, demonstrated integration, or independent adoption of HarnessRouter’s cross-harness API. It therefore does not change the case from an architecturally relevant but unvalidated launch.
2026-08-20T11:23:18Z
evidence attached: hn.story.49373059 — The artifact directly bears on whether coding agents can use interchangeable model backends, a core claim of the open harness-as-backend case.
2026-08-19T22:34:26Z
No new evidence has appeared after the launch discussion, so the case remains an architecturally relevant but unvalidated bet rather than an emerging adoption signal. Revisit only when an independent implementation, user, benchmark, or harness-provider integration appears.
2026-08-17T22:30:45Z
The refreshed discussion still adds no independent user, implementation, benchmark, or provider participation; it remains repetitive amplification of the launch and its unresolved differentiation question. The architecture is relevant to Scott, but the adoption hypothesis is still untested.
2026-08-17T21:37:22Z
The refreshed comments remain repetitive founder-led explanation plus the same unresolved differentiation challenge; they add no independent implementation, adoption, or proof that a canonical cross-harness API delivers practical value. The case remains architecturally relevant but commercially and technically unvalidated.
2026-08-17T20:35:09Z
The refreshed discussion surfaces a concrete differentiation challenge—why a router is needed when individual harnesses expose reusable servers or cores—but the only response remains first-party. This sharpens the product-risk question without adding independent adoption or validation.
2026-08-17T19:45:25Z
The slight engagement increase adds no independent validation, implementation evidence, or consequential participant. HarnessRouter remains a highly relevant but unproven first-party launch rather than evidence that harness-as-backend integration is gaining adoption.
2026-08-17T19:34:57Z
grounded: known/high — HarnessRouter directly implements Scott's agent-native computing architecture: it provides the machine-native abstraction layer between agent runtimes and consu
2026-08-17T19:29:44Z
origin walked (codex/luna, conf 0.94): anchor hn.story.49335595 -> echo.blog.532ab429b0 by Richard Song
2026-08-17T19:27:20Z
case created — First-party GitHub release of a unified harness API, but zero comments and low engagement indicate no independent validation yet.