Already 6 months old when the scout caught it.
Advisor pattern
What it is
The advisor pattern (also called the advisor strategy or two-model pattern) splits an agentic task across two models with different jobs: a cheap, fast "executor" model drives the task end to end — calling tools, writing output, taking actions — while a stronger, more expensive "advisor" model never touches tools or users and only gives guidance when the executor hits a decision it can't handle confidently. The executor decides when to consult the advisor, distilling the problem, evidence, and options rather than forwarding the whole transcript. This inverts the older orchestrator pattern, where a large model delegates subtasks to smaller workers — here the small model stays in the driver's seat and escalates upward instead.
Why it's emerging now
As agentic workflows got longer-horizon and more expensive to run at frontier-model quality, teams needed a way to keep most of a task on a cheap model without losing the accuracy a stronger model provides. Anthropic formalized this as a first-class API primitive (the "advisor" tool) in April 2026, exposing it as a one-line addition to a Messages API call rather than something developers had to hand-roll. Independent practitioner coverage, an open-source "advisor swarm" implementation in the swarms.world library, and a community-built Claude Code plugin ("Advisor Opus") show the pattern being picked up and reimplemented outside Anthropic's own docs — distinct from merely reading about the feature. It also surfaced independently as a named design pattern at the AI Engineer Europe 2026 conference, where it was grouped with academic "advisor model" work as a now-common orchestration idiom.
How it relates to what you know
It sits next to the existing "Managed agents" and "Harness engineering" entries in the ledger as another orchestration-layer idiom, but it specifically answers the cost/quality tradeoff that multi-agent orchestration and mixture-of-experts address at the model-architecture level — instead of routing to different experts inside one model, or spinning up a swarm of peer subagents, it pairs exactly two models in a fixed executor/advisor hierarchy where only the executor acts. It's a narrower, cheaper cousin of full multi-agent orchestration: one upward escalation path instead of a graph of peers.