Surviving Anthropic Rate Limits via Model Orchestration

Cole Medin is tackling a very specific and timely pain point for teams pushing AI coding agents to their limits: Anthropic's degraded rate limits. His 'Dark Factory' experiment—a codebase that evolves itself from GitHub issue to deployed PR without human review—is a high-token consumer, making it the perfect testbed for a pragmatic workaround. He explicitly acknowledges that full autonomy isn't production-ready, but uses it to stress-test model orchestration.

The core engineering decision today is splitting the workflow by intelligence tier, not just random task. He's shifting the architecture of his Archon workflows (YAML-based agentic pipelines) to use Claude Opus strictly for the 'planning' node and Kimi K2.6 as the workhorse for everything else (research, implementation, validation). This directly addresses the cost and availability constraints engineering leaders are facing: the most powerful models are heavily rate-limited on subscription plans. The debate he highlights is genuinely practical—is superior reasoning more valuable during the specification/planning phase, or when rubber meets the road in code generation? He leans toward planning, arguing a perfect spec reduces downstream rework.

The implementation detail that matters is the painful provider switching. He can’t just use the Anthropic SDK globally and route it to different endpoints; mixing Opus and Kimi requires abandoning the Claude SDK 'hijack' that was routing everything to Kimi, and instead using a separate open-source agent ('Pi') to call Kimi's API, while simultaneously re-authenticating the native Claude toolchain against a real Anthropic subscription. This exposes the brittle state of current multi-model orchestration—there’s no clean unified interface.

He also confirms the adversarial validation pattern is critical without human review: a second agent reviews the PR with zero context of the implementation conversation, seeing only the issue and the final diff. This 'separation of concerns' for AI reviewers prevents bias creeping in from the builder agent’s thought process. The stream is a live look at the messy reality of stitching together subscription APIs, managing tokens, and building fault tolerance into autonomous pipelines, not a polished product demo.