The agent cost wars are being won before the first invoice

The bill that never lands

LEASH is a training-free framework for computer-use agents that routes work between a weak model and a strong one based on latent uncertainty — and the uncertainty estimator reuses activations from action generation, so it needs no extra forward pass and no dedicated monitor training. On OSWorld it outperformed random escalation by 6.55 to 11.94 points. The money line: Qwen3.5-35B-A3B with LEASH hit 73.66% task success at $0.4117 per task, against 72.90% for always-on Claude Opus 5 — a 19.2% cost reduction with a small accuracy gain. A shadow observer keeps watching the strong agent after handoff. The cheapest frontier call is the one you never make, and LEASH is learning which ones those are. The routing decision itself requires no extra forward pass, so the overhead of the savings is near zero — the uncertainty signal comes free with the action. m:…4293

The intern that does the clicking

JevAgentHarness attacks the same bill from the tool-use layer. LLM agent harnesses currently call a large LLM at every step; Jev delegates tool choices to small decision models using fitted confidence thresholds — and the threshold is load-bearing, because delegating without it lowered task success by 19.0 points. On tasks similar to its training data, 4B models trained this way made agents 81.1 to 89.0% cheaper and up to 2.5 times faster. The researchers also released AgentDecision-4B, an open 4B model trained on agent error states with Jev's answers as labels, plus AgentDecisionBench to test decision models from isolated questions up to control of a running agent. The click is no longer a frontier-model decision. The approach also held up off-distribution: on tasks unlike its training data, the agent solved 91.5% of them with AgentDecision-4B and 82.0% with Jev. m:…3784

The edge keeps the easy steps

EdgeWise moves the split to the network boundary. A local small language model handles the run, and a step-level reinforcement learning method selectively delegates complex steps to cloud frontier models — allocating the LLM to only a small subset of key steps while achieving performance comparable to full LLM execution. The result: up to 62% lower inference cost on SWE-Rebench and 34% on Terminal-Bench 2.1, demonstrated on a local workstation. No additional models are required at inference time. The edge does the routine; the cloud gets the hard steps. The step-level RL jointly optimizes task solving and delegation with fine-grained credit assignment, so the delegation policy improves as the agent does. m:…7996

Draft cheap, verify once — and keep the spread

SpecAgent formalizes the pattern as a protocol: a small model drafts tool-use steps — executing read-only tool calls as it goes so drafts stay grounded in real observations — and the strong model verifies drafted segments in a single prefill pass, committing the longest prefix it judges semantically valid and repairing only the first rejected step. On τ²-Bench the success rate stayed within one point of the strong-only model while memory traffic and wall time roughly halved; in MCP-Atlas deployments, strong-model token generation fell to at most half. Beyond τ²-Bench, the paper reports SpecAgent recovering 44–75% of the weak-strong quality gap on AutomationBench and VitaBench, and closing 53–75% of it in MCP-Atlas deployments. This suggests the emerging margin structure for agent products: the strong model becomes the reviewer, not the author. For founders, the decision is architectural — build the draft-verify-escalate loop now, because the token bill is being engineered down before it ever lands, and whoever captures that spread keeps it. m:…5173