The PGE harness: why Anthropic spends 15x more on Claude and still calls it cheap

The mistake most multi-agent demos make

The popular framing is that multi-agent systems work because more agents are smarter than one. That is the wrong abstraction.

Multi-agent demos that ship in production do not win on aggregate IQ. They win on two architectural choices: context isolation (the Evaluator does not see the Generator's reasoning, so it cannot inherit the Generator's blind spots) and deterministic routing (the harness code, not the model, picks which tool subset each role sees). When teams skip both and just spin up three Claude calls that pass chat history around, they get the worst of every world — higher cost, slower iteration, and the same blind spots they started with.

The four patterns below are what the Anthropic demo actually demonstrates, once you strip out the marketing.

Four patterns the PGE harness retires

Pattern 1. Decomposition is the work; the model is not.

The Planner does not "think harder." It runs on xhigh effort (the new mid-tier between high and max introduced with Claude Opus 4.7 on April 16, 2026) and produces 5-10 user stories with explicit acceptance criteria. The Generator never sees the high-level request. It receives one story at a time, with a clear definition of done. That is the whole point: the Generator's failure modes shrink when the surface area of its task shrinks. The model did not get better. The task got smaller.

Pattern 2. The Evaluator must be structurally isolated.

A single agent grading its own code suffers from self-evaluation leniency. Anthropic's reference setup runs the Evaluator as a separate agent with its own system prompt — "Senior QA Lead / Adversarial Auditor" — and gives it Playwright MCP to interact with the live application. The Evaluator returns structured JSON (status, critique, playwright_logs, recommended_fix), not prose. The pedantic rule that ships with the demo: "If a criterion is 99% complete, it is a FAIL." That is not theatre; it is the only way to prevent the loop from terminating on optimism.

Pattern 3. The handoff is a file, not a conversation.

Passing chat history between agents is the agent telephone game. Every handoff loses a little intent and adds a little noise; by iteration five the Evaluator is judging work against a goal the Planner never set. Anthropic's documented practice is the Shared Structured Artifact — a single evolving markdown spec (often AGENTS.md or CLAUDE.md at the repo root) that all three agents read and update. The spec is durable across retries, survives context-window rollover, and gives the Evaluator a stable target. The Planner writes; the Generator annotates; the Evaluator critiques in place.

Pattern 4. Task Budgets are not a nice-to-have.

The new Task Budgets API caps the entire loop, not just one call. A multi-agent harness without a hard ceiling can quietly burn 15x-19x more tokens on the wrong problem and you would not notice until the monthly bill. The April 16 release notes ship Task Budgets alongside Adaptive Thinking (dynamic reasoning-token allocation per sub-task), so the harness can spend deeply on the Planner's decomposition and shallowly on the Generator's third unit test. Without budgets, the economics never close. With them, the 15x-19x premium is bounded and decidable.

The nuanced point

Multi-agent harnesses are not a universal upgrade. The benchmark data is unambiguous: below 12 distinct tool-domains, a single agent with prompt caching outperforms the PGE harness on cost and matches it on quality. The jump from 32.1% to 84.0% success on >12 tool-domain tasks is real, but it is also a threshold function. Spin up three agents for a task that touches only Git and a database, and you have paid the harness tax for nothing.

The decision is therefore architectural, not aspirational. Ask whether the task crosses tool-domain boundaries (browser + DB + shell + API + Git, say), whether the agent will run for more than a handful of iterations, and whether the cost of an undetected failure is higher than the cost of three agents in a loop. If all three are yes, the PGE harness pays for itself. If any is no, a single Claude Opus 4.7 call with the right system prompt is the more honest answer.

Where does your current agent setup fail first — the model, the tools, or the handoff?

The honest answer for most teams is the handoff. The model is fine. The tools are fine. But the agents are passing chat transcripts around like a relay race where each runner forgets the baton, and by lap three the goal has quietly drifted. The PGE harness is not interesting because it has three agents. It is interesting because it forces the question of what survives between them — and the answer, written down, on disk, in markdown, is the boring architectural choice that makes the 15x premium worth paying.