A Known Workflow Needed Fewer Model Decisions

When an agent already knows the required sequence of work, should the model still decide every tool call? WTK tested that question by comparing model-directed tool use with a governed capability that carried the known sequence itself.

The result supports a bounded answer: in this experiment, fewer model decisions produced more reliable completion and used less model work.

Question

The tested workflow converted place names into verified coordinates and then calculated straight-line distance. The dependency order was already known: resolve the origin, resolve each destination, verify locality, and calculate distance only when the required coordinates exist.

WTK asked whether deterministically scheduling that known segment could reduce tool-selection failures without concealing empty results, failed attempts, clarification requests, or skipped downstream work.

Experiment

The experiment used two matched arms. In the first, the model selected and coordinated the individual location and distance tools. In the second, the model invoked one governed capability whose internal steps and dependencies were fixed.

Both arms used the same Nova Pro model binding, goals, candidate inputs, locality constraints, temperature, call limits, deterministic tool responses, and scorer. Four fixtures covered clean completion, verified-address fallback, wrong locality, and an empty origin. Each fixture ran three times per arm, producing 12 runs per arm and 24 runs in total. Arm order alternated between repetitions.

Every run counted. There were no hidden agent retries, discarded provider failures, or best-of selection. Only the tool surface changed. The model calls were live, while the constituent tool responses were fixed locally so the comparison isolated orchestration rather than network or search drift.

Result

Measure	Model-directed tools	Governed fixed workflow
Goal accomplishment	9 of 12	12 of 12
Factual honesty	11 of 12	12 of 12
Output conformance	11 of 12	12 of 12
Interaction correctness	11 of 12	12 of 12
Provider calls	37	23
Model tool calls	37	15
Input and output tokens	54,347	38,879

The governed arm won three matched outcomes and tied nine. It had no losses. Provider calls fell by 37.8 percent, model tool calls fell by 59.5 percent, and total tokens fell by 28.5 percent.

The difference was not merely shorter output. One model-directed run changed a valid receipt identifier into a plausible but invalid value. Another stopped after an empty primary lookup instead of trying the declared fallback. The governed workflow retained exact receipts, executed the fallback, halted for clarification when locality was wrong, and skipped dependent work when the origin could not be established.

Bounded finding

When a workflow segment and its dependency order were already known, representing that segment as a governed deterministic capability improved accomplishment in this matched sample while reducing model-visible orchestration work.

This does not argue for making all agent reasoning deterministic. The model still handled the goal and the surrounding judgment. WTK moved only the known mechanical sequence into inspectable structure and preserved the evidence from every internal step.

Boundary

This was one place-distance workflow, one model binding, one local execution target, four fixtures, and fixed tool responses. It does not establish the same improvement for other models, harnesses, open-ended research, live data providers, larger samples, or workflows whose correct sequence is not already known.

The measured latency is descriptive, not a provider benchmark. No general cost claim is supported.

Builder impact

Agent builders should distinguish reasoning problems from already-known dependency chains. Leaving a known chain entirely to model tool selection creates additional opportunities to omit a required call, corrupt a reference, or continue after evidence has become insufficient.

Moving that chain into structure is useful only if the structure remains honest. It must retain child receipts, represent empty and failed attempts, stop for clarification when required, and refuse to fabricate downstream success.

Next experiment

WTK should repeat the matched protocol with another model, another harness, a different known workflow, and live tool responses. A larger replication should test whether the reliability difference survives those changes or narrows to this specific execution environment.