{"slug": "cadence-v8-4-a-multi-model-coding-harness-where-claude-writes-codex-reviews-and", "title": "Cadence v8.4: a multi-model coding harness where Claude writes, Codex reviews, and Bugbot triages", "summary": "Cadence v8.4, an open-source coding harness, now routes different software development lifecycle roles to different AI models, allowing Claude to write code, Codex to review it, and Bugbot to triage issues. The developer shipped the update this week, arguing that using a single model for authoring, reviewing, and triaging introduces blind spots and overconfidence, similar to letting a junior engineer self-review their own pull request. The harness supports 16+ providers and risk-tiered review depth, and has been dogfooded to ship itself across multiple versions.", "body_md": "Claude writes. Codex reviews. Bugbot triages. Gemini sits on the council.\n\nCadence is an open-source coding harness — same category as Devin, Cursor agents, Aider, and Goose, but with one structural bet: different SDLC roles run on different models. The model that writes the code is not the model that reviews it, and neither of them is the model that triages the bug report. You get to pick the model per phase, and the orchestration is just a YAML file you can edit.\n\nI shipped v8.4.0 this week. Here's why the multi-model angle matters and what the harness actually does.\n\n**The single-model problem**\n\nMost coding agents (including the well-funded ones) run the whole SDLC on one model. The same weights write the diff, review the diff, and decide whether the bug report is real. That's the equivalent of letting a junior engineer self-review their own PR and also decide which Sentry alerts to dismiss.\n\nA model that wrote subtly broken code is statistically the worst model to catch the bug in it. They share the same blind spots: the same fencepost errors, the same favorite anti-patterns, the same overconfidence on the parts they got wrong. Tier-1 human teams don't review like this. The author writes. Someone else reviews. A third person triages production incidents. Different people, different cognitive frames, different incentives.\n\nCadence treats LLMs the same way. The author and the reviewer should be different models, ideally from different families, ideally trained on different data mixes.\n\n**What Cadence does**\n\nCadence is a pipeline of rewireable skills:\n\nbrainstorm -> spec -> plan -> implement -> migrate -> validate -> PR -> review -> bugbot -> merge\n\nEach phase is a skill: a markdown contract plus a small amount of harness glue. Each phase can route to a different provider. The defaults I run:\n\nThe interesting failure modes the multi-model setup catches: Claude over-mocks tests, Codex catches it. Codex over-fixates on imaginary race conditions, Claude flags the false positive. Gemini disagrees with both about a schema decision, council resolves it. Nothing about this is magical, it's just three independent passes by models that don't share weights.\n\n**Risk-tiered review depth**\n\nEvery spec declares a risk tier in its frontmatter. The harness reads it and picks how many review passes to run before merge:\n\nThis is honest engineering ROI. A typo fix doesn't need three rounds of GPT-5.5. A migration that touches production tables does.\n\n**Concurrent multi-PR dispatch**\n\nv7.11.0 shipped worktree isolation. You can run N specs in parallel, each in its own git worktree, each in its own Cadence session, each producing its own PR. I've shipped 4 PRs concurrently in one sitting without any cross-contamination. The harness handles the worktree lifecycle; you just point it at specs.\n\n**16+ providers**\n\nThe provider adapter layer covers: Anthropic, OpenAI, Google (Gemini), Groq, Ollama, AWS Bedrock, Azure OpenAI, Cohere, Mistral, DeepSeek, Together, Fireworks, Perplexity, OpenRouter, xAI Grok, and any OpenAI-compatible endpoint you point it at. Every phase can be overridden in a profile:\n\nphases:\n\nimplement:\n\nprovider: anthropic\n\nmodel: claude-opus-4-7\n\nreview:\n\nprovider: openai\n\nmodel: gpt-5.5\n\ncouncil:\n\nprovider: google\n\nmodel: gemini-3.5-pro\n\nbugbot_triage:\n\nprovider: anthropic\n\nmodel: claude-haiku-4-7\n\nSwap providers without touching skill code. If a new model lands tomorrow, change one line.\n\n**Receipts**\n\nCadence ships itself. Most v7.x and all of v8.x went through its own pipeline: spec written by Cadence, implemented by Cadence, reviewed by Cadence, bugbot-triaged by Cadence, merged by Cadence. You can browse the merged PRs here:\n\n[https://github.com/axledbetter/cadence/pulls?q=is%3Apr+is%3Amerged](https://github.com/axledbetter/cadence/pulls?q=is%3Apr+is%3Amerged)\n\nThe dogfood is the test suite.\n\nClaude Code as a distribution surface.\n\nOne housekeeping note: Cadence's skills load through Anthropic's Claude Code CLI as one of the distribution surfaces. You don't need to be a Claude Code user to care about the harness: the orchestration, the multi-model routing, the worktree dispatch, the bugbot triage are all provider-agnostic. Claude Code happens to be a convenient host for the skill loader. That's the whole relationship.\n\nTry it\n\nnpm install -g @delegance/cadence@latest\n\ncadence autopilot examples/specs/node-cli.md\n\nRepo: [https://github.com/axledbetter/cadence](https://github.com/axledbetter/cadence)\n\nLicense and footprint\n\nMIT. Local-first. Runs on your machine with your API keys. No hosted agent, no telemetry, no per-seat subscription, no \"contact sales.\" If your employer says you can run npm install and call provider APIs from your laptop, you can run Cadence.\n\nIf you've been waiting for an open-source coding harness that doesn't pretend a single model is good at every job, this is the one I built. Pull requests, issues, and skeptical takes all welcome.", "url": "https://wpnews.pro/news/cadence-v8-4-a-multi-model-coding-harness-where-claude-writes-codex-reviews-and", "canonical_source": "https://dev.to/axledbetter/cadence-v84-a-multi-model-coding-harness-where-claude-writes-codex-reviews-and-bugbot-triages-1d73", "published_at": "2026-05-27 13:00:00+00:00", "updated_at": "2026-05-27 13:11:06.180974+00:00", "lang": "en", "topics": ["ai-agents", "large-language-models", "ai-tools", "ai-products", "ai-infrastructure"], "entities": ["Claude", "Codex", "Bugbot", "Gemini", "Cadence", "Devin", "Cursor", "Aider"], "alternates": {"html": "https://wpnews.pro/news/cadence-v8-4-a-multi-model-coding-harness-where-claude-writes-codex-reviews-and", "markdown": "https://wpnews.pro/news/cadence-v8-4-a-multi-model-coding-harness-where-claude-writes-codex-reviews-and.md", "text": "https://wpnews.pro/news/cadence-v8-4-a-multi-model-coding-harness-where-claude-writes-codex-reviews-and.txt", "jsonld": "https://wpnews.pro/news/cadence-v8-4-a-multi-model-coding-harness-where-claude-writes-codex-reviews-and.jsonld"}}