Anthropic published a companion white paper on agent architecture alongside a detailed blog post this month—a pairing that positions the company as a source of opinionated architectural guidance for engineering teams, not just frontier model capability.
The March 5 post on the Claude blog distills three workflow patterns into production-ready blueprints: sequential, parallel, and evaluator-optimizer. The patterns are framed as composable building blocks rather than rigid templates, using a manufacturing metaphor—an assembly line where individual workers retain decision-making authority but the overall process is architected in advance.
Sequential is the recommended default for tasks with hard dependencies—multi-stage data pipelines, draft-review-polish content cycles, document extraction followed by schema validation. Each agent waits for the previous step to complete, which adds latency but improves accuracy by keeping each agent focused on a scoped subtask. Anthropic's practical advice: start with a single agent and only introduce sequential steps when a lone agent demonstrably can't handle the full task reliably.
The parallel pattern applies when tasks are logically independent and speed is the constraint. Borrowing a fan-out/fan-in model from distributed systems, multiple agents execute concurrently—useful for multi-dimensional evaluations, simultaneous code review across modules, or running identical tasks with varied parameters. Compute and token costs scale with concurrency, and the post flags result aggregation as a design decision worth resolving upfront.
The evaluator-optimizer pattern closes an iterative loop: a generator agent produces output, an evaluator scores or critiques it, and the cycle repeats until a quality threshold is met. Anthropic targets this at domains where output quality matters more than inference cost—technical documentation, customer communications, code generation against strict standards.
Together, the blog post and white paper are Anthropic's clearest attempt yet to standardize vocabulary and decision frameworks for engineering teams building production agent systems.