Add another agent when it performs a necessary, checkable job that the simpler workflow cannot do as well within your constraints. Before expanding a recurring workflow, compare the delivered result, total run cost, waiting time, and human correction effort against a simpler version. Keep required oversight in both versions.
For a digital or IT leader, the useful question is specific: what does the proposed planner, reviewer, or specialist contribute that would otherwise be missing? A convincing role description is a hypothesis to test.
What this week’s research found
Zettl and Gosmar’s October 25 preprint reports 2,430 runs across 135 coding tasks using Claude Opus 5.5 and Haiku 5.5. Its all-strong-model planner–coder–reviewer pipeline cost 3.22 times and took 3.24 times as long as one strong-model call, without a demonstrated accuracy gain. Study design and Table 3.
Replacing the planner and reviewer with the cheaper model reduced cost to 0.39 times the all-strong pipeline. Against the single strong call, however, that mixed pipeline still cost 1.27 times as much and lost 8.1 percentage points of pass@1, the share of runs passing all task tests. These are different comparison baselines. Results.
The measured protocol accounted for 0.11% of pipeline wall time on one host. That is not a cross-network estimate. The study also distinguishes an oracle model selector, which knows outcomes afterward, from its implemented router. An oracle result is not deployable routing performance. Protocol and routing results.
This is a preprint covering two same-vendor models and likely training-exposed coding benchmarks. The cheaper reviewer produced the graded final artifact; planning and review effects are not isolated. The authors participate in the initiative developing the evaluated protocol. AMF has not reproduced the study. Validity limits.
Give each additional agent an acceptance test
The following is AMF’s proposed decision procedure, not a validated measurement standard.
Start with the job that needs doing. For a hypothetical weekly operations report, the required result might be a reconciled summary, a list of exceptions, and links to the supporting records. A person must approve external distribution.
Now describe the proposed specialist’s contribution as an observable result:
- A reconciliation agent finds discrepancies against the supplied records.
- A source-checking agent identifies unsupported statements and supplies evidence.
- A reviewer detects a defined class of errors before the report reaches its owner.
“Review the report carefully” is too vague for this comparison. “Check every reported total against the supplied export and identify mismatches” gives the evaluator something to verify.
Do not count agreement as a contribution by itself. If a second agent approves a wrong total, the workflow still has a wrong total. If it flags a valid total and creates unnecessary rework, record that too.
Compare the workflow with and without one optional stage
Use the same inputs, acceptance criteria, permissions, and required human review. Remove only the stage whose value you are testing. Preserve any control required by policy, privacy, or the consequences of an error.
For the reporting example:
- Baseline: One agent prepares the report; a person checks it using the agreed acceptance criteria.
- Candidate: The same preparation step plus a reconciliation specialist; the person applies the same criteria.
- Decision: Determine whether the specialist catches consequential errors, reduces human correction, or improves delivery enough to justify its measured overhead.
Include ordinary inputs and troublesome cases: missing records, inconsistent totals, and an exception that changes the conclusion. Repeat cases rather than keeping each system’s best attempt. Keep failed and unfinished runs visible. A small trial can expose a bad design; it cannot establish reliability for every future case.
This differs from comparing agent harnesses. Here the change under examination is a specific stage inside the workflow. Avoid simultaneously replacing the model, rewriting every prompt, and adding a reviewer, then crediting the reviewer for the outcome.
Record what the extra stage actually changes
Keep a compact record for each attempt:
- Result: Accepted, rejected, blocked, or unfinished, with the reason.
- Additional contribution: Which error the specialist caught or which required output it supplied.
- Damage introduced: New mistakes, unnecessary changes, or false alarms.
- Resource use: Available model and tool charges, retries, and elapsed time.
- Human effort: Time spent reviewing, correcting, and resolving disagreements.
Use measured charges where available. If the product exposes only a subscription allowance, record the available usage measure and its limits rather than inventing a per-run price. The AI pilot cost guide separates the cost categories; the human-work log helps keep intervention visible.
Report quality and expense together. A low cost per accepted result can hide an unacceptable rejection rate. Conversely, a stage that increases machine expense may still be useful if it demonstrably reduces consequential errors or human rework.
Download a cost-comparison sheet
Use the blank extra-agent cost-comparison CSV for a baseline and a candidate with one optional stage. The hypothetical worked example CSV shows how to record an additional agent’s contribution, false alarms, model and tool charges, retries, elapsed time, and human effort. The shared worksheet instructions and field definitions explain the calculations. Import the CSV as a comma-separated UTF-8 file; it is an editable log, not an automatic calculator.
Match the task cases and repetitions across conditions. Keep required oversight, the preparation model, inputs, and acceptance criteria unchanged. Include every agent and retry in the candidate’s costs and elapsed time. Judge the submitted result before material human correction and record later repair time separately.
Every number in the worked example is invented. No trial was run. Across four matched reporting cases, the fictional reconciliation specialist catches a totals mismatch, creates a false alarm on a correct report, misses an unsupported narrative claim, and cannot replace a missing input.
- Baseline: One accepted, two rejected, and one blocked run; USD 0.85 total model-plus-tool cost, 310 elapsed run seconds, and 29 active human minutes.
- Candidate: Two accepted, one rejected, and one blocked run; USD 1.32 total cost, 455 elapsed run seconds, and 23 human minutes.
- Cost per accepted result: USD 0.85 for the baseline and USD 0.66 for the candidate, retaining all failed and blocked runs in the cost total.
This hypothetical candidate has a lower cost per accepted result but higher total machine expense and waiting time. One rejected case remains. Those tradeoffs and the false alarm belong in the decision; the example is not evidence to keep or remove a real stage.
Report unavailable costs as unknown, not zero. Cost per accepted result is undefined with no accepted output or incomplete cost coverage. Keep currencies consistent, state exclusions, and report subscription usage separately when no run price is available. Human minutes remain separate from charges and elapsed seconds.
Make a keep, narrow, or remove decision
Keep the stage when it supplies a needed result or catches an important error class at an acceptable cost.
Narrow the stage when its useful contribution is limited. In the hypothetical report, a reconciliation specialist may need only the totals, source rows, and exception rules. Asking it to rewrite the whole report expands the work being evaluated and creates another opportunity to change correct content.
Remove the optional stage when it mostly repeats the first agent’s work without a demonstrated benefit. Keep the checks that establish the final output’s acceptability.
Escalate selected cases when the evidence suggests only difficult inputs need extra attention. Test the selection rule itself: an inexpensive workflow that fails to recognize difficult cases has not solved the problem.
These are evaluation choices, not instructions to weaken an existing approval or security boundary.
Limitations and next action
This guide does not establish that single-agent systems are always better, that reviewers are unnecessary, or that business reporting behaves like code generation. It proposes a bounded comparison rather than a product ranking. The reporting scenario is hypothetical, and no savings or error reductions have been measured by AMF.
Choose one optional stage in a recurring workflow. Write down the exact defect or missing output it should address. If that contribution cannot be described clearly, clarify the role before adding another agent. If it can, compare the simpler workflow with the candidate and make a decision from the delivered work.
Provenance
Prepared for Agent Model Fit with AI-assisted primary-source inspection, analysis, and drafting on October 30, 2026. Research findings are attributed to Zettl and Gosmar’s version 1 preprint. The business example and decision procedure are AMF proposals. No independent replication, product certification, human review, or measured business outcome is claimed.
Update, November 4, 2026: Added downloadable blank and hypothetical worked-example CSVs with field definitions. The original research summary and conclusion are unchanged. The worksheets were prepared with AI assistance; their example values are invented, not measured results.