Claude Opus 4.8’s Dynamic Workflow Enables Hundreds of Parallel Subagents

Anthropic released Claude Opus 4.8 after a series of roughly monthly minor updates (4.5 → 4.6 → 4.7 → 4.8). The most visible improvement is the model’s honesty: it is about four times less likely to claim a defect‑free code when errors exist, and it more often admits uncertainty instead of fabricating answers.

Using a public harness, the author benchmarked Opus 4.8 against 4.7, GPT‑5.5, and Gemini 3.1 Pro. On the SWE‑Bench Pro coding suite Opus 4.8 scored 69.2 % versus 64.3 % for 4.7, while GPT‑5.5 lagged at 58.6 %. In the Agentic terminal coding benchmark (Terminal‑Bench 2.1) Opus 4.8 achieved 74.6 % but was outperformed by GPT‑5.5’s 78.2 %.

The headline feature is the dynamic workflow, a research‑preview capability added to Claude Code (requires version v2.1.154+). Instead of the traditional sub‑agent/skill approach where Claude decides the next step, the workflow moves the planning logic into a generated JavaScript script. The script can launch dozens to hundreds of parallel sub‑agents, collect their outputs in variables, and return a single converged answer to the user.

Two ways trigger a workflow: (1) include the word “workflow” in the prompt, which makes Claude generate a workflow script; (2) open an ultracode file (/effort ultracode) where Claude performs high‑intensity reasoning to decide whether a workflow is worthwhile and may launch multiple stages automatically.

A flagship example from Anthropic shows Bun’s author using a dynamic workflow to port ~750 k lines of code from Zig to Rust, achieving 99.8 % test‑suite pass and completing the merge in 11 days—a task that would normally take months.

The author’s own test involved a medical‑imaging AI research request. The workflow split the task into five phases with a total of 111 agents: 1 Scope, 6 Search, 28 Fetch, 75 Verify, and 1 Synthesize. In the Verify phase, each of 25 key conclusions was checked by three independent agents, resulting in 75 verification agents. Many verification agents failed to invoke the StructuredOutput tool, causing 17 of the 25 conclusions to be marked as “killed”. Opus 4.8, however, paused before summarising and explicitly distinguished confirmed, unverified but likely true, and clearly refuted statements, demonstrating its honesty.

Running workflows consumes a large number of tokens and quickly hits the 5‑hour usage limit for most accounts, limiting large‑scale experiments. Nevertheless, the author finds the dynamic workflow more exciting than the model upgrade itself and encourages readers to try it on suitable long‑running tasks.

Finally, Anthropic hinted at an upcoming “Mythos” model, described as the strongest model on Earth, slated for release in the coming weeks.