Before Adding Another AI Agent, Measure the Handoff

A services firm designs a proposal workflow with separate AI agents for understanding the request, checking delivery capacity, and preparing the scope. Each specialist performs well in a demonstration.

During a complete run, the capacity agent identifies a staffing dependency. Its summary reaches the scope agent without the condition attached. The final proposal looks polished, but the team must reopen the analysis before it can approve the work.

This hypothetical example is a coordination problem. Adding another specialist might help, or it might create another place for a critical detail to disappear. The decision needs evidence from the complete workflow.

New guidance puts agent count back on the table

On October 22, Anthropic published guidance on folding subagents into a main agent loop. The company reports that some enterprise implementations benefited from consolidating routing, planning, and domain work. It also identifies reasons to retain subagents, including independent parallel work, verification, and separate permission boundaries.

These are the vendor's observations and recommendations. They are not an independent comparison establishing the best design for every business, and we have not reproduced the reported customer results.

The broader tradeoff predates this update. Anthropic's December 2024 guidance on effective agents discusses adding complexity when it improves task performance, with cost and latency as tradeoffs. That is useful background, rather than a new release.

For business leaders, our recommendation is to ask what each agent boundary contributes to an accepted outcome. A larger team of agents is an architectural choice with an operating cost.

Give each handoff a business purpose

Map where responsibility passes from one agent to another. For each boundary, record the question being answered, the evidence sent across, and the result the receiving agent needs.

In the proposal example, the capacity result should distinguish confirmed availability from availability that depends on a hiring decision. A short summary that drops that distinction is not a successful handoff, even if the specialist's original answer was correct.

Keep source references and unresolved conditions attached to the result. Make it possible to inspect the evidence behind the final recommendation without replaying the entire conversation.

Then ask why the work is separate. Perhaps it needs a different tool, an independent review, or a different access boundary. Perhaps the split exists because separate departments built separate components. The first set of reasons may justify the coordination cost. The organizational history alone does not establish a benefit for the person receiving the proposal.

Price the coordination work

Count more than the model calls that produce the final answer. Include routing, specialist preparation, summaries, repeated lookups, and repairs after incomplete handoffs.

Measure elapsed time from the original request to a result the business can accept. Keep human review and correction effort visible. A faster specialist is of limited value if the complete proposal spends longer in clarification.

Also assign the maintenance work. Someone must keep the shared result format consistent when one specialist changes. Someone must investigate failures that cross team boundaries. Record that effort alongside infrastructure and model charges when comparing designs.

Use the same acceptance definition for each candidate. As with measuring AI value at the end of the workflow, a comparison becomes misleading if one design receives credit for a draft while another must deliver reviewed work.

Compare one boundary before redesigning the system

Select a tightly connected pair of steps for a controlled comparison. In the proposal workflow, that might be capacity assessment and scope preparation, because the second depends directly on the first.

Use approved test records and the same business questions. Keep the existing design as the baseline. Compare a candidate that preserves the relevant context in one working process, while retaining the required validation and approval steps.

Inspect individual cases, including those that improve and those that regress. Include a conditional answer, missing evidence, conflicting instructions, and a request outside the allowed scope. Run variable cases more than once so a single good demonstration does not decide the rollout.

Record whether the final proposal preserves the staffing dependency, how long review takes, and whether the reviewer can trace the evidence. Set the decision criteria before examining the results. The candidate should earn its place through the business outcome, not the neatness of its diagram.

Keep the boundaries that protect the business

Consolidating reasoning must not quietly combine permissions. If one component may read staffing data and another may prepare customer-facing material, review what the combined design can now access and disclose.

Keep enforcement in the appropriate systems and tools. Preserve the authorized reviewer and the required approval before a proposal leaves the organization. Recheck the agent permission controls whenever responsibilities or tool access change.

Retain a separate specialist when the evidence supports it. Independent verification may justify another step even when it adds delay. Parallel research may help when its outputs can be combined without constant clarification. The important question is whether the separation performs a necessary job under the actual workload.

Release the simpler or stronger design deliberately

Give one owner responsibility for the complete workflow, including the comparison results and unresolved failures. Roll out a change to a bounded workload first, with a clear way to return to the previous version if quality falls.

Monitor correction effort and recurring handoff failures after release. A design that works on the initial sample can meet different demands as users and business rules change.

If your AI workflow has accumulated specialists, talk with Revival Group about testing one handoff against its business purpose, operating cost, and permission boundary before adding more agents.