Anthropic introduced dynamic workflows for Claude Managed Agents on 9 October 2026, adding a beta feature that lets an agent write a program to coordinate other agents across a larger task. The company points to reviewing hundreds of documents as one use case. Its server executes the resulting workflow in the background. Anthropic release notes
For businesses with backlogs of pages or documents to check, the announcement raises a practical question: can a coordinated AI review produce a more useful, verifiable worklist than the current process? That is the question worth testing before connecting the system to live business operations.
What changes with dynamic workflows
Developers enable the feature in an agent's multi-agent configuration and tell it when to use a workflow. It remains in beta under the existing Managed Agents beta header. This is a capability for applications built with Managed Agents, rather than a promise that every Claude chat will automatically process a document backlog. Release details
A workflow can split work into phases, run agents in parallel and pass their findings into subsequent steps. It can also repeat a step or handle a failure. The current documentation lists 64 simultaneously working threads per run, a figure Anthropic says can change, and a separate limit of 1,000 agents started over the run's lifetime. The latter is a cumulative limit. Workflow documentation
The business opportunity is in tasks that naturally divide into smaller checks. Think of a website owner reviewing product information across a catalogue, or an operations team comparing a set of approved documents against a consistent checklist. A useful output would identify each issue, show its evidence and leave a reviewer with a clear next action.
Completion needs a quality check
One important detail in Anthropic's documentation deserves attention: a workflow marked completed can still contain failed work. Organisations also retain their existing model rate limits. Teams therefore need to inspect individual results and failures before accepting the overall output. Workflow behaviour
A review dashboard should show how many items were assigned, successfully checked, skipped and left unresolved. Every finding should link back to the relevant page or document section. Missing evidence should remain visible as a gap, rather than disappearing into a confident summary.
That distinction affects procurement, too. Ask a supplier to demonstrate what happens when a page cannot be retrieved, a file is unreadable or two checks disagree. A polished final report says little about whether the underlying review covered everything it was supposed to cover.
The hourly rate is only part of the cost
Anthropic lists Managed Agents runtime at US$0.08 per session-hour while a session is running. Model tokens are billed separately, and tools can add charges; web searches inside a session cost US$10 per 1,000 searches. The standard Batch API discount does not apply to Managed Agents sessions. Managed Agents pricing
A business case should include integration, testing and human review alongside usage charges. Measure the cost per accepted finding or successfully reviewed item. An inexpensive run that creates hours of checking and rework may offer little value.
Set a conservative pilot allowance and leave headroom. Anthropic notes that in-flight model requests can take a workflow beyond its session budget before it pauses. Budget behaviour
A sensible first pilot for a South African business
South Africa appears on Anthropic's supported-country list. That supports country eligibility, but does not establish a particular account's beta access or guarantee processing capacity. Confirm those details before committing to delivery dates. Supported countries
A practical starting point would be a read-only review of 30 to 50 public pages on a business's own website. This is a proposed pilot, not a reported Symaxx Digital test. Choose a narrow question, such as whether service descriptions contain inconsistent opening hours, locations or product details.
Prepare the expected answers for a small sample first. Then require a source link for every finding and compare the output with a human review. Record missed issues, false alarms, review time and total cost. Include an unavailable page deliberately to see whether the system reports the gap accurately.
Keep publishing and record changes behind human approval. Begin with public or synthetic material; using confidential documents should wait until the organisation has checked its permissions, security requirements and provider terms. Agree who can stop the pilot and what happens to incomplete work.
What to decide next
Dynamic workflows give developers another way to organise large review jobs. For buyers, the next decision is which tightly scoped backlog has clear success criteria and a manageable cost of error.
Start there. Expand only when the pilot shows that the output is reliable, the evidence is easy to check and the total effort is lower than the process it replaces.