# FastDelegate

By Hao-Wei Lee

Model recommendations for a predefined task, balancing fit evidence, pricing, context, tools, quota, and recorded outcomes.

Source: https://github.com/egginsect/fast-delegate

Canonical: https://hwl.dev/work/fast-delegate
Last content review: 2026-10-07

## Which worker can complete this task with the tools and budget available?

FastDelegate turns one task and the active harness’s available model selectors into a ranked recommendation list. The caller still chooses the worker, dispatches the task, and verifies the result.

## Design decisions

### Check the constraints first

Runtime, context, tool support, known quota, and a known price below the lead are safeguards before recommendation.

### Keep judgment with the caller

Fit scores are advisory evidence, not success probabilities. Every recommendation requires lead review; the router executes no workers.

## Evidence and boundaries

Evidence below summarizes the public project documentation, not an independent audit. Source README: https://github.com/egginsect/fast-delegate/blob/main/README.md

### An explicit recommendation boundary

Status: Implemented

The public README documents routing, an explicitly labeled heuristic fallback, and outcome recording after caller verification.

Source record: Public README: Recommend a model, Jev and fallback, Discovery and feedback
Citation: https://hwl.dev/work/fast-delegate#recommendation-boundary

## Current limitations

- One predefined task per recommendation request; task decomposition and dispatch belong to the caller.
- Missing or stale evidence remains unknown. Fit scores are not measured success probabilities.
- Subscription and local deployments use shadow cost, not claimed dollar savings.

## Synthetic walkthrough

Illustrative data; not a live service or measured result.

### 1. Describe a bounded task

An illustrative routing request, not a benchmark or live model call.

Input: A task with acceptance criteria, the lead model, and available worker selectors.

Decision: Supply the actual context and tool requirements for this task.

Output: A routing request grounded in the active harness.

### 2. Review the recommendations

Compare eligible candidates using the available evidence.

Input: Ranked recommendations with model identity, estimated cost, and fit evidence.

Decision: Choose a credible full-completion candidate; do not treat rank or price as automatic acceptance.

Output: A worker choice for the caller to dispatch.

### 3. Verify and record

The caller owns the handoff and checks the finished work.

Input: The worker’s deliverable and the original acceptance criteria.

Decision: Accept or reject based on verification, then record the observed outcome.

Output: Outcome evidence that can inform later recommendations.

A public prototype landing page is not available yet.

Contact: https://www.linkedin.com/in/hao-wei-lee-b7ba2b37/