Hybrid-model harness for coding is the standard now

Cognition's Devin Fusion Sets the Template

Cognition shipped Devin Fusion, a hybrid-model coding harness claiming a 35% cost reduction while keeping fable-level intelligence quality. The architecture splits work across models — an expensive planner for hard decisions, cheaper models for bounded subtasks. This sakan and fugu-style routing is now the pattern every coding agent is copying.

Mid-Session Routing Solves the Cache Problem

Walden Yan and Jerry Liu of Llama Index broke down sidekick and mid-session routing publicly this week. A smaller, faster model jumps in mid-session for repetitive work while the big model stays available for critical calls. Jerry specifically flagged the cache efficiency advantage — bouncing between models and losing context burns tokens and latency. Delegate to cheap models, but preserve cache locality and context continuity.

Agents Are Becoming Programmable Control Planes

Sydney Runkel, Harrison Chase, and the LangChain team are converging on the same idea: agents should write orchestration code, not just call tools. This shifts agents from tool-using chatbot to something closer to a Kubernetes-style scheduler for large task fan out. The main agent reasons about the task, then writes the script to distribute it — a fundamentally different architecture than a ReAct loop.

Retrieval Harnesses Beat Grep

Llama Index shipped a retrieval harness combining semantic search, grep, file listing, and file reading in one agent loop. Max Peeperkorn pushed back on the grep is all you need crowd — real-world retrieval is messier than text search, and agents need multiple retrieval modalities working together.

Cheap Trajectory Judges Fix Eval Costs

Harrison Chase announced TraceJudge, a model for detecting trajectory errors in agent runs at roughly 1/100 the cost of closed models. Finding where an agent went wrong is now affordable enough to use in the inner loop.