A planner-and-worker design assigns difficult decisions to one model and much of the execution to another. It is attractive when a task needs a strong initial plan but then contains many straightforward steps. The economic question is whether the cheaper execution outweighs the additional context, consultations and coordination required to keep both models aligned.
Google’s November 12 Antigravity Orchestrator announcement and Anthropic’s advisor tool offer two concrete versions of this idea. They are useful starting points for an evaluation, not a guarantee that any pair will be cheaper or more reliable. Current documentation and prices below were checked on November 17, 2026; the worked calculation is explicitly hypothetical, with no new benchmark or paid inference run.
- 01Find a real division of workPairing is most plausible when difficult decisions are intermittent and the worker can complete the intervening steps without constant rescue.
- 02Count every modelInclude planning, execution, repeated context, advice and failed attempts. The worker’s visible token total may be only part of the bill.
- 03Define when to reconsiderUse new evidence, repeated verification failures or an invalid assumption as reasons to replan. Routine activity does not need a fresh consultation.
- 04Compare accepted resultsMeasure total spend per accepted task alongside quality and elapsed time. A cheaper attempt that frequently needs repair may be a worse operating choice.
01 — Workload fitStart with the shape of the task
A good candidate has an uneven distribution of difficulty. A software change might require careful dependency analysis followed by a series of bounded edits and checks. A research workflow might need an experienced judgement about scope before collecting a set of well-defined facts. In both cases, the planner’s contribution is valuable if it makes the remaining work easier to execute correctly.
Other tasks offer little room for that division. A short answer may finish before a handoff could add value. A task that requires difficult judgement at every step may keep returning to the planner until the worker becomes an expensive intermediary. If a single model already performs the job quickly and reliably at an acceptable cost, adding another model needs a specific reason beyond architectural interest.
Begin with examples from the actual queue. Mark where the task became difficult, what evidence resolved the uncertainty and how much routine work followed. Our Gemini Hydrogen, Argon and Carbon comparison helps frame model selection, but model strengths alone do not determine the best workflow. The task’s structure, tools and acceptance criteria decide whether a useful division of labour exists.
| Task pattern | Pairing hypothesis | Reason to stay simple |
|---|---|---|
| Occasional difficult decisions | Planner guides a long execution phase | The handoff cannot preserve essential context |
| Repeated bounded operations | Worker handles most steps | The worker needs frequent correction |
| Short response | Little planning overhead is needed | A second call may dominate the task |
| Continuous hard reasoning | A planner remains heavily involved | One capable model may be more direct |
02 — Two patternsDistinguish dynamic handoffs from an advisor call
Google’s Antigravity model documentation describes Orchestrator as a larger model planning and reviewing while Hydrogen implements. The planner starts in a read-only phase, and the models can hand control back when new evidence or repeated verification failures require a revised plan. This is a changing division of work within a task, rather than routing the whole request once at the beginning.
Anthropic’s advisor tool lets an executor request strategic guidance from a separately invoked advisor. The advisor receives the full transcript, returns advice and the executor continues. The two patterns should not be treated as identical products: one coordinates execution phases, while the other adds a consultation within an executor-led flow. Use only the model pairs and configuration supported by the chosen implementation.
Google reports that Orchestrator achieved near-Carbon task success on Terminal-Bench 4.1 while consuming 24% less quota and completing 19% faster on average in its November 12 announcement. Those are vendor-reported results, checked November 17. Quota is not an API-dollar bill, and a benchmark average does not predict every repository’s outcome. Treat the result as motivation to test your own task distribution, not a savings figure to insert into a business case.
Change who owns the next phase
A coordination layer decides when work should return to planning and when execution can resume.
Keep the executor moving
The executor receives guidance during its task and remains responsible for continuing the work.
Judge the complete trajectory
Measure the result, cost and time of the whole workflow rather than treating each model call as a separate success.
03 — Handoff designGive the worker a plan it can verify
A useful plan describes the objective, relevant constraints, the current evidence and the conditions that would invalidate the approach. It should tell the worker what completion means. A long narrative about the planner’s reasoning is less useful than a clear statement of which files, facts or dependencies matter and which checks must pass. The handoff should reduce ambiguity rather than require the worker to reconstruct the original investigation.
For a hypothetical application change, the planner might identify the affected flow, the expected behaviour and the existing checks that demonstrate it. The worker then implements that bounded change and reports concrete results. If the implementation reveals a dependency the plan did not account for, that is new evidence for replanning. If a routine edit succeeds and the relevant check passes, repeating the entire planning phase adds little information.
The multi-agent orchestration playbook covers the broader coordination problem. Keep ownership explicit when adapting the pattern: one agent should know whether it is executing, reviewing or deciding how to proceed. Overlapping roles can produce duplicated work or incompatible changes. A compact handoff with a clear acceptance condition often matters more than adding another model to an already ambiguous process.
State the objective, evidence, constraints, completion checks and reasons to return for a decision. Preserve the facts the worker needs to recognise when the plan no longer fits.
04 — Cost ledgerCount advice, repeated context and retries
Create a ledger for the whole task. Record input and output usage for every model, including billed reasoning where applicable, and account for cache treatment, tool charges and repeated attempts. A handoff may send the same source material to another model. A later consultation may include a longer history. Those tokens are part of the architecture’s cost even when the worker’s final answer is short.
Anthropic’s documentation makes one accounting detail explicit: top-level usage totals cover the executor, while advisor calls are separately billed at advisor rates. Use usage.iterations to include both. Later executor iterations can also process the growing history again. An integration that records only the top-level total can therefore understate the cost of the consultation pattern. Verify the billing fields for the actual product and API route in use.
Keep token spend separate from elapsed time and human effort before combining them into a business decision. A slower task may still be acceptable in a background queue; a slightly higher API bill may be worthwhile if it materially reduces expert repair time. Conversely, an apparently cheap run can be expensive if someone must untangle its output. Report the dimensions separately so the team can see which trade-off the architecture is making.
| Cost component | Why it appears | What to record |
|---|---|---|
| Planner or advisor | Initial decisions and later consultations | Model, usage and call count |
| Worker | Implementation and routine steps | Aggregate request usage |
| Repeated context | History crosses calls or model boundaries | Billed input after cache treatment |
| Failure and recovery | Retries, replanning and human repair | Spend and effort before acceptance |
05 — Worked exampleSee how a small overhead can erase the saving
For an illustrative direct-API calculation, use Google’s standard rates per million tokens: Carbon at $2 input and $10 output through November 30, 2026, and Hydrogen at $1.50 input and $7.50 output. Carbon’s listed rates rise to $4 and $20 on December 1. The pricing page was checked November 17. Output includes billed thinking; this example excludes caching, tools, infrastructure and human time.
Assume a Carbon-only task uses 100,000 aggregate input tokens and 20,000 output tokens, costing $0.40. A hypothetical split using Carbon for 20,000 input and 4,000 output tokens, plus Hydrogen for 80,000 input and 16,000 output tokens, costs $0.32. These token quantities are chosen assumptions, not measured equivalent performance. The comparison only becomes commercially meaningful if both designs produce an accepted result.
Now add 30,000 Hydrogen input tokens for duplicated handoff context: the mixed cost becomes $0.365. Add one further Carbon consultation of 20,000 input and 4,000 output tokens, and it reaches $0.445, above the original $0.40. The lesson is not that pairing fails. It is that a plausible saving can be sensitive to ordinary coordination overhead. Use aggregate billed tokens, including repeated context, rather than counting each source document only once.
| Hypothetical design | Carbon input / output | Hydrogen input / output | Token cost |
|---|---|---|---|
| Carbon only | 100k / 20k | 0 / 0 | $0.400 |
| Initial split | 20k / 4k | 80k / 16k | $0.320 |
| Split plus context | 20k / 4k | 110k / 16k | $0.365 |
| Plus another consultation | 40k / 8k | 110k / 16k | $0.445 |
06 — Escalation rulesReturn to planning when the evidence changes
Define a small set of reasons to escalate: a material assumption is contradicted, the same relevant check repeatedly fails, a required dependency is missing or the next action exceeds the worker’s authority. These triggers describe a change in the task state. Calling the planner after every successful action adds expense without necessarily improving the decision. Never escalating can be equally wasteful when the worker continues along a failed approach.
Budget controls should match the implementation. In Anthropic’s Messages API, the advisor tool’s max_uses applies per turn, which can span requests; its own output cap is separate from the executor’s. Managed Agents does not expose those same advisor cap options. A limit attached to the worker alone should not be assumed to cap the entire system. Review both call-count and output controls where the product provides them.
Also define what happens when a limit is reached. The worker might stop with a clear unresolved issue, return a partial result for review or continue only with actions already authorised by the accepted plan. It should not quietly replace an unavailable decision with a guess merely to finish. The guide to agentic workflow anti-patterns explains why explicit stopping and recovery paths are part of reliable automation, not just cost management.
Will another consultation resolve a concrete uncertainty that changes the next action? If not, the system may be adding discussion rather than progress.
07 — Verification boundaryKeep advice separate from evidence and authority
A second model is not automatically an independent verifier. If both models see the same incomplete transcript, they may share the same missing fact. Agreement is useful only in relation to the evidence available and the checks performed. For a software task, inspect the relevant files and test results. For a research task, verify the source. For a business action, establish who has authority to approve it.
The planner can help decide which checks matter, while the executor gathers their results. A later review should examine the resulting artifact and the evidence of completion, not simply accept the worker’s statement that it followed the plan. If a check does not test the required behaviour, passing it offers limited assurance. Make the acceptance condition concrete enough that another person could understand what remains uncertain.
Keep permissions attached to the task rather than the model’s apparent confidence. A strong planner does not turn an unapproved external action into an approved one. A fast worker does not need broader credentials merely because it performs more steps. Separate the ability to recommend an operation from the ability to execute it, and preserve a human decision where the workflow requires one. This prevents a cost optimisation from quietly changing the authority of the system.
08 — Pilot decisionCompare cost per accepted task against a real baseline
Choose a representative set of tasks and define acceptance before comparing architectures. Run the existing single-model workflow and the proposed pair under comparable tool access, task scope and review criteria. Include straightforward cases as well as difficult ones, because a queue composed only of complex tasks can overstate the value of constant planning. Keep model versions and relevant settings visible in the result.
Calculate cost per accepted task as total spend, including failed attempts and retries, divided by the number of accepted tasks. For success rate, divide accepted tasks by all attempts. Report elapsed time and human correction effort separately. A design that saves tokens but increases failed tasks may lose its apparent advantage once the complete delivery process is counted. Likewise, a higher token bill may still be a reasonable trade if it produces more dependable results.
Our AI transformation services connect that evaluation to a specific business workflow. Adopt the pair when the evidence shows a useful improvement at an acceptable quality and operating cost. Keep the single-model path when it remains the better fit, and preserve a fallback for tasks that repeatedly need the stronger model. The architecture should follow the observed work, with a clear explanation of why the extra coordination earns its place.
The pair should improve an outcome you care about—accepted-task cost, quality, elapsed time or human effort—without hiding a material deterioration in the others.
Make the second model earn its place
Planner-and-worker systems can be effective when a small amount of stronger judgement makes a large amount of execution more reliable. Their economics depend on the complete trajectory: both models, repeated context, consultations and recovery. Start with a clear handoff and a capable baseline, then choose the design that produces accepted results at the best practical cost for your workload.