Claude Opus 4.8 makes agent work bigger, so reviews need smaller slices

Anthropic's new Claude Opus 4.8 is easy to read as another model upgrade. Faster, better at code, stronger at agentic tasks. The more useful part for normal teams sits one layer down: Claude Code is starting to treat bigger jobs as workflows, not just one long chat.

Source: Introducing Claude Opus 4.8

What actually changed

Anthropic introduced Claude Opus 4.8 on May 28, 2026, and describes it as stronger at coding, agentic tasks, reasoning, and practical knowledge work. The same release also points to better self-checking: Opus 4.8 should more often flag uncertainty and notice flaws in its own code.

The more practical change is in the Claude Code 2.1.154 changelog: dynamic workflows. You can ask Claude to create a workflow, and Claude can then coordinate tens or hundreds of background agents. Runs are visible through /workflows. Version 2.1.156 followed shortly after with a fix for Opus 4.8-related API errors, so I would treat this as a current signal, not something to roll out everywhere without review on day one.

Source: Claude Code CHANGELOG.md

Why this matters for small teams

Dynamic workflows in Claude Code mean an AI tool can split up a larger assignment and let several agent runs work in the background. That is not only useful for large codebases. The same pattern can help a Nordic team review an integration, clean up documentation, compare system settings, or prepare a migration.

But the review has to get smaller, not bigger. If Claude can start more parallel tracks, the human needs to see the work in narrow packets: what would change, which files or systems are touched, what evidence exists, and where approval is required.

That is a good Tool Forge question: do not build a magic AI button. Build a routine where the agent has the right access, works in clear slices, and leaves receipts before the next step.

Try it like this

Choose one recurring but bounded task. For example: prepare a test plan for our customer portal, compare our integrations against the docs, or find old onboarding text that needs updating. Give Claude read access to the relevant files or documents. Hold back write access until the plan has been reviewed.

Try this prompt this week:

Read the README, test setup, relevant config files, and any agent or MCP settings.
Propose a dynamic workflow for: [task].
Split the work into 3-5 reviewable packets.
Show which files or systems each packet may touch, what evidence you want to collect, and where I should approve the next step.
Do not start the workflow yet. Start only after I choose one packet.

A good answer is not the biggest plan. A good answer makes the work easier to approve or reject.