Anthropic says Claude Code can write JavaScript harnesses that split work across agents, route models by task and add reviewers to curb drift in long software jobs.
Anthropic said June 15 that Claude Code can create JavaScript execution harnesses for Dynamic Workflows, a Claude Code feature that coordinates groups of agents on large engineering tasks.
The company framed the approach as an orchestration layer for work that outgrows a single prompt, single model choice or single context window. Anthropic’s post builds on its June 1 Dynamic Workflows announcement for Claude Code, which targets parallel agent coordination in software projects.
Anthropic says the model can generate harness code that delegates work, assigns agent roles, checks results and decides how long a workflow should run. A team could use one agent to inspect an API boundary, another to write tests and a reviewer agent to challenge both outputs before a developer accepts the patch.
The technical bet sits in coordination. Long agent runs often fail because the model quits early, grades its own answer too favorably or drifts away from the original goal. Anthropic calls those failure modes agentic laziness, self-preferential bias and goal drift.
Dynamic Workflows try to reduce those failures through separation of work. Anthropic describes fan-out-and-synthesize workflows, adversarial verification, tournament-style runs and classifier systems that route jobs by complexity. In a fan-out pattern, Claude Code can split a task into parallel subtasks, collect results and merge the useful parts. In an adversarial pattern, a reviewer agent attacks another agent’s conclusion before the developer sees the final answer.
Model routing adds another systems trade-off. Anthropic says teams can assign cheaper models to routine stages and reserve stronger models for tasks that need deeper reasoning. That pattern mirrors distributed system design: spend scarce capacity on the bottleneck, keep common paths cheap and add checks around the places that can corrupt the result.
Developers still have to pay for the extra coordination. More agents mean more tokens, more intermediate state and more places to inspect when the run fails. Some developers welcomed the control. Others questioned whether today’s reliability gains justify the cost. In one Reddit discussion, a user said the approach can burn tokens before it produces enough value.
The design points toward a broader shift in AI engineering. Model quality still matters, but teams now spend more effort on orchestration, verification and routing. A single model answer gives you one path through a problem. A harness gives you competing paths, reviewers and a record of intermediate decisions.
That record matters for software teams. A developer can inspect which agent changed a file, which agent rejected a claim and which model handled each stage. Without that audit trail, multi-agent work becomes a black box with a larger bill.
Anthropic has positioned Dynamic Workflows beside other Claude Code automation features, including managed agents and routines. The company’s developer material sits in the Anthropic documentation and the Claude Code product page.
The approach will appeal to teams that already run code review, CI and staged deployment as verification systems. It will frustrate teams that expect an agent swarm to replace those controls. Claude Code can coordinate more workers, but engineers still need to define success, inspect failures and decide when the harness has earned its cost.
Comments
Please log in or register to join the discussion