AgentDecision-4B Makes Agent Loops 89% Cheaper

// 1h agoMODEL RELEASE

AgentDecision-4B Makes Agent Loops 89% Cheaper

AgentDecision-4B is a LoRA-tuned Qwen3-4B adapter that handles bounded tool, argument, and completion choices inside agent harnesses, deferring uncertain states to larger language models. The paper reports 81.1–89.0% lower costs and up to 2.5× faster execution on familiar tasks, plus 91.5% success on shifted tasks. [Paper](https://arxiv.org/abs/2610.21412)

// ANALYSIS

The compelling idea is not simply using a smaller model, but teaching it where agent trajectories actually go wrong and calibrating when it should hand control back to the LLM.

–Post-mistake trajectory training targets the distribution shift that breaks many standalone tool-choice classifiers.
–Per-choice confidence thresholds are essential: uncalibrated delegation reportedly reduced task success by 19 points.
–The adapter handles bounded decisions efficiently while preserving frontier-model calls for planning, recovery, and difficult reasoning.
–Results are promising but remain author-reported, and developers must recalibrate thresholds for their own harnesses and workloads.
–The project points toward hybrid agent stacks where small local models manage routine control flow and large models handle semantic complexity. [Repository](https://github.com/JevAgentHarness/JevAgentHarness)

// TAGS

agentdecision-4bsmall-llmopen-weightsfine-tuningagenttool-useinference

DISCOVERED

1h ago

2026-10-18

PUBLISHED

1h ago

2026-10-18

RELEVANCE

9/ 10

AUTHOR

Discover AI

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE9m ago

AI Job Search Turns Claude Code Into Career Agent

[AI Job Search](https://github.com/MadsLorentzen/ai-job-search) turns Claude Code into a local job-application pipeline that profiles candidates, finds and scores roles, tailors CVs and cover letters, and prepares interviews. Its creator says the workflow produced 69 tailored applications, 20 first-round interviews, and one signed contract before he became an AI engineer.

RESEARCH1h ago

JevAgentHarness routes agent decisions with confidence gates

JevAgentHarness is an open-source proxy that routes routine agent decisions to fast models while escalating uncertain steps to stronger LLMs. Its AgentDecisionBench research reports major cost and latency reductions using per-choice confidence thresholds and the open AgentDecision-4B model.

OPEN SOURCE1h ago

git-ai brings server-centric Git to agents

git-ai is an MIT-licensed, open-source Git platform built for concurrent coding agents. It centralizes history, sparse data access, landing, and configurable conflict resolution on a server while retaining Git underneath. ([GitHub repository](https://github.com/feder-cr/git-ai))