- Who it helps
- Developers and researchers evaluating local agents for extended coding or terminal work.
- Availability
- The paper and evaluation details are public. The project page still marks code as forthcoming; no ready-to-run implementation was verified.
The decision happens during the task
The authors train a Qwen3.5-9B agent to choose when to delegate. They evaluate coding and terminal tasks with GPT-5.5 or Claude Opus 4.8 supplying selected responses. This makes delegation part of the learned behavior.
Read the cost boundary
The project reports task-success and cloud-usage tradeoffs. Its dollar figures cover cloud API usage, including its judge where used. Local inference costs are excluded. A smaller cloud bill therefore does not establish a lower total operating cost.
A comparison worth preparing
Keep the task, tool permissions and acceptance tests identical across a local-only run, a cloud-only run and a delegation run. Record completed tasks, elapsed time, local compute and API charges. Count failed attempts and retries in the same ledger. This is a proposed comparison; it has not been run here.
What you can do
- Use the paper to define a comparison now. Wait for a versioned implementation before planning a reproducible EdgeWise deployment.
Where the evidence stops
Reported results belong to the authors’ evaluation setup. Code availability, independent reproduction and whole-system costs remain open.
Sources read. Sources checked October 18, 2026. No product workflow or benchmark was executed.
Sources & further reading
Paper · paper
EdgeWise paper ↗
Selected for direct evidence about the subject, its availability or the limits of the claim.
EdgeWise authors · page checked 2026-10-18
Web · paper
EdgeWise project and evaluation details ↗
Selected for direct evidence about the subject, its availability or the limits of the claim.
NVIDIA Research · page checked 2026-10-18
Coverage note: Independent hands-on comparisons and current cross-platform discussion are incomplete in this bounded sample. X public search required login.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
EdgeWise studies a practical split: a small local agent does most of a task and asks a larger cloud model for help on selected steps.