A Harness for Every Task: Claude Code's Architectural Guide to Dynamic Workflows

Why Single-Context Agents Hit a Wall

Anthropic's June 2 guide to dynamic workflows opens with a frank diagnosis of where agentic coding breaks down. Three failure modes emerge when a single agent context tackles sufficiently large or complex work:

Agentic laziness: the agent stops making progress before the task is actually complete, typically on work that requires sustained effort across many files or decisions.

Self-preferential bias: the agent evaluates its own output favorably rather than subjecting it to genuine scrutiny, producing results that pass a superficial check but miss deeper issues.

Goal drift: over many turns, the agent gradually loses fidelity to the original objective, substituting plausible-but-wrong subgoals.

Dynamic workflows address all three by breaking work into isolated subagent contexts, each with a focused objective and no access to the distorting history of prior turns.

Seven Orchestration Patterns

The guide catalogs seven workflow architectures Claude Code can generate on demand:

Classify-and-act

Route incoming tasks by type before processing. Useful for support queue triage, bug categorization, or routing code to specialized reviewers.

Fan-out-and-synthesize

Distribute work across parallel subagents, then merge and reconcile results. The canonical use case is a codebase-wide audit where each agent handles a subset of files.

Adversarial verification

Spawn a dedicated challenger agent that independently verifies the output of a primary agent. Forces genuine scrutiny rather than self-evaluation.

Generate-and-filter

Create a broad set of options or candidates, then apply a separate filtering pass. Useful for generating test cases, refactoring candidates, or API design alternatives.

Tournament

Run multiple agents on the same task and compare outputs competitively. Particularly effective for decisions where "best" is hard to specify up front.

Loop until done

Repeat a workflow iteration until a condition is met. Useful for fuzzing, retry-with-backlog, or tasks with non-deterministic completion.

Memory and rule adherence

Maintain explicit state across iterations, enforcing constraints that individual agents might drift from over time.

When Not to Use Workflows

The guide is notable for its explicit caution: dynamic workflows consume significantly more tokens than single-context runs. Routine coding tasks (writing a function, fixing a bug, reviewing a PR) do not benefit from orchestration overhead. Anthropic recommends reaching for workflows only when the task genuinely requires more coordination than one conversation can sustain, or when the orchestration itself (as a reusable script) is part of the deliverable.

Availability

Dynamic workflows are in research preview for Max, Team, and Enterprise plans via the Claude Code CLI, Desktop app, and VS Code extension, as well as the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry.