Multi‑model coding workflow gains ground
- GitHub and model makers are pushing a new coding pattern: one model plans, another executes, and a separate reviewer checks the work. - GitHub added “Rubber Duck” to Copilot CLI on April 6, pairing Claude Sonnet 4.6 with GPT-5.4 to review plans and code. - The shift tracks wider support for parallel agents and bring-your-own models in coding tools. (github.blog)
A new coding habit is taking shape in 2026: developers are starting with one model, handing execution to another, then asking a third pass to challenge the result. (github.blog) (openai.com) (anthropic.com) GitHub made that pattern explicit on April 6, when it introduced “Rubber Duck” in experimental mode for Copilot CLI. The feature uses a second model from a different family to review an agent’s plan and output before work moves ahead. (github.blog) GitHub said a Claude Sonnet 4.6 session paired with Rubber Duck running GPT-5.4 closed 74.7% of the performance gap between Sonnet and Claude Opus 4.6 on hard SWE-Bench Pro tasks. The company said the gains were strongest on problems spanning three or more files and more than 70 steps. (github.blog) That is a more formal version of a workflow many developers have been describing in public: use a reasoning-heavy model to map the job, let a coding agent make targeted edits, then run an adversarial review that looks for missed assumptions, edge cases, and security issues. (github.blog) (docs.github.com) The basic idea is simple. A planning model acts like a tech lead writing the spec, an execution model acts like the engineer changing files and running tests, and the review pass acts like a skeptical code reviewer trying to break the plan. (github.blog) (openai.com) (anthropic.com) The tooling is moving in the same direction. OpenAI says Codex is “designed for multi-agent workflows,” with built-in worktrees and cloud environments for parallel work across projects. (openai.com) Anthropic is also framing coding as orchestration, not single-shot prompting. On its Claude Code product page, Anthropic says its own engineers now focus on architecture, product thinking, and “continuous orchestration,” including managing multiple agents in parallel. (anthropic.com) GitHub has added the plumbing to make model-mixing easier. On April 7, it said Copilot CLI could connect to Anthropic, Azure OpenAI, OpenAI-compatible endpoints, or local models, letting developers keep the same terminal workflow while swapping providers. (github.blog) GitHub also rolled out `/fleet` on April 1, a Copilot CLI command that breaks a task into independent work items and dispatches multiple sub-agents in parallel. The orchestrator decides what can run at the same time, polls for completion, and synthesizes the final output. (github.blog) The result is that “multi-model workflow” is becoming less of a creator tip and more of a product category. The common thread across GitHub, OpenAI, and Anthropic is not one best model, but a stack of specialized roles. (github.blog) (openai.com) (anthropic.com)
