Anthropic’s multi-agent harness for long-running app development
Anthropic Engineering says Prithvi Rajasekaran moved from a single long-running coding harness to a multi-agent setup with planner and evaluator roles, using graded criteria to handle taste, correctness, and spec quality.
What happened
In an Anthropic Engineering post, Prithvi Rajasekaran says he was working on two related problems: getting Claude to produce higher-quality frontend designs and getting it to build complete applications without human intervention. He describes extending earlier frontend-design and long-running coding harness work, then taking inspiration from GANs to build a multi-agent structure with generator and evaluator roles. The evaluator was designed to grade outputs reliably, including subjective taste, by turning judgments into concrete criteria. Rajasekaran also says he added a planner agent that turns a 1–4 sentence prompt into a fuller product spec, while staying at the level of product context and high-level technical design to avoid cascading implementation errors.
Why it matters
FACT: Anthropic Engineering says Prithvi Rajasekaran was trying to improve Claude’s frontend design output and its ability to build complete applications without human intervention, then moved toward a generator-evaluator setup with a planner that expands brief prompts into fuller specs. INFERENCE: The practical decision is whether app-building agents need explicit planning and grading layers, not just a stronger base model.
Engineering impact
INFERENCE: For teams building long-running coding systems, the useful constraint is to isolate load-bearing harness parts and remove them one at a time, because the post says a radical simplification failed to preserve performance and later complexity reduction was done methodically. That implies evaluation should track which orchestration components actually stabilize outcomes under changing model capability.
Leadership impact
EDITORIAL_INTERPRETATION: The article supports a management choice to invest in process and architecture, not only model selection, when agentic products must meet taste, correctness, and usability targets. It also suggests product leaders should treat spec quality as a controllable design variable, while recognizing that the source presents one company-reported method rather than an externally validated standard.
Editorial judgment
This changes a real engineering decision: whether to add a planner/evaluator layer around agentic app-building rather than keep a single harness. The article is strongest as a company-reported method description; it does not establish independent reproduction, so readers should treat the design as a plausible pattern, not a validated benchmark winner.
Watch next
FACT: Rajasekaran says later iterations were motivated in part by Opus 4.6, which Anthropic described as needing less scaffolding for longer agentic tasks, larger codebases, code review, debugging, and long-context retrieval. INFERENCE: The next decision to watch is whether stronger base models keep shrinking the need for planner and evaluator scaffolding, or merely change where it is still valuable.
Limitations & uncertainty
The supplied source is partial and company-authored, so the analysis is limited to Rajasekaran’s described methods and Anthropic’s own framing. It does not establish independent reproduction, benchmark comparability across unknown setups, or general performance claims beyond the reported harness iterations.