Stop calling tools one by one. Code mode cuts the token bill by up to 90 percent
Three new studies point the same way: how an agent calls tools matters more than which model does the calling. The cheapest wins come from the interface, not the model.
- Code mode cuts token usage by 70 to 90 percent against plain MCP while significantly boosting utility. Self-reported m:…7402
- Declaring return-shape hints in code mode stubs swings success by up to 39.5 points for lower-performing agents. Self-reported m:…7402
- Endpoint scoring fails to distinguish agents that retrieve evidence efficiently from those that read most of the environment, per DetourBench. Self-reported m:…7404
- Skill Transfer improves agent pass rates on weak models by repairing structural gaps in execution trajectories. Self-reported m:…7391
The interface beats the model
The cheapest way to make an agent better at tools may be to stop making it call them one at a time. A new study comparing code mode with plain MCP finds that code mode substantially outperforms plain MCP, cutting token usage by 70 to 90 percent while significantly boosting utility. Declaring return-shape hints in code mode stubs swings success by up to 39.5 points for lower-performing agents. The paper also tests the obvious next step, a decision layer in front of code mode, and reports that the hypothesis that it beats code mode alone is not confirmed. m:…7402
That result reframes where the budget goes. Teams have been buying bigger models to fix tool use; this paper says the interface is doing more of the work. A stub that tells the agent what shape the answer comes back in is worth more than another layer of orchestration, at least for the agents that struggle most. m:…7402
Nudge, lightly
A second study looks at what happens when you coach the agent while it works. ToolPulse, a middleware layer, can identify intervention opportunities using state-based rules, semantic sequence analysis, and interaction-feature similarity. The study finds concise reminders more effective than elaborate assistance in eliciting substantive behavioral responses from agents. But it also finds that behavioral adjustment does not guarantee improved aggregate repair outcomes: the agent changes what it does without necessarily fixing more. m:…7392
The gap between behaving differently and succeeding more is the warning. A nudge that makes the trace look cleaner is not the same as a nudge that repairs the task, and teams that measure the trace instead of the outcome will fool themselves. m:…7392
Measure the path
The measurement problem is the subject of a third paper. DetourBench, a trajectory-level evaluation framework, argues that endpoint scoring fails to distinguish between agents that retrieve evidence efficiently and those that read most of the environment. It computes a full-information minimum action count and exact residual work in partially observable environments, and its conditional attainment metrics can re-rank model performance compared with standard success-based metrics. m:…7404
A fourth paper attacks the same efficiency question from the model side. Skill Transfer, a framework for adapting agent skills from strong to weak agents, improves agent pass rates on weak models by repairing structural gaps in execution trajectories. It uses Graph Edit Distance to localize where strong and weak agent trajectories diverge, and the adapted skills improve performance on strong agents too and transfer to structurally related, unseen tasks. m:…7391
Taken together, the four papers say the same thing: the agent's bill and its score both depend on the path it takes, not just the destination. Price the trajectory, hint the return shape, nudge briefly, and distrust any success rate that never looked at how the agent got there. m:…7402 m:…7404
The practical consequence is a change in what teams instrument. If endpoint scoring cannot tell an efficient agent from one that reads most of the environment, then success rates alone will keep rewarding the wasteful agent. DetourBench's answer is to score the trajectory itself, counting the minimum actions a fully informed agent would need and the work left over, so that a team can see whether a cheaper interface actually made the agent smarter or merely luckier. m:…7404
What is still unresolved
- Whether code mode's token savings hold for agents that already perform well, or only for lower-performing ones
- Whether concise reminders keep working once agents adapt to them
- Whether DetourBench-style trajectory scoring changes model rankings on production tasks, not just benchmarks
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.