Quantum Dice Trinity Challenge 2026 / Team 48 (Solo)

Probabilistic Scheduling for Agentic Coding Workflows

1 / 6

Problem

Multi-agent coding systems need routing before they spend tokens.

Planners, researchers, coders, testers, reviewers, and tool runners share real limits: token budgets, context windows, API quota, tool calls, and deadline pressure.

Naive routing can waste capacity and put conflicting work on the same agent before the workflow engine notices.

Jobs coding, tests, review
Agents specialized capabilities
Budgets tokens, context, tools
Risks conflicts and overload

Scheduler

The model chooses job-to-agent assignments and reports violations honestly.

Input workflow batch

jobs, agents, capabilities, dependencies, conflicts, budgets

QUBO x[job, agent] = 1

assignment cost, penalties, capacity slack, communication terms

Output assignment plus diagnostics

cost, feasibility, resource usage, violation types

The implemented MVP is assignment-first and explainable; time-indexed deadline variables stay as a future extension.

ORBIT implementation

ORBIT samples the QUBO; decoding and repair make the result schedulable.

Raw, decoded, and repaired results stay separate. A repaired feasible assignment is not presented as raw ORBIT feasibility.

Captured evidence

The benchmark story is useful because it includes both wins and failure modes.

Toy workflow -

ORBIT repaired matches the exact-small optimum.

Medium workflow -

ORBIT repaired is feasible; random repair is lower at -.

Congested workflow -

No tested solver found a feasible assignment.

The strongest claim is not broad ORBIT superiority. It is a real probabilistic sampler path with clear repair and feasibility reporting.

User value

This sits before an AI workflow engine, where mistakes are still cheap.

Route specialist work

Match coding, research, testing, and review jobs to agents with the right capabilities.

Estimate resource pressure

Expose token, context, and tool-capacity overload before dispatch.

Keep failure visible

Report infeasible outputs and remaining violations instead of hiding them behind a score.

Compare honestly

Show probabilistic sampling beside classical baselines on the same scenarios.

Interactive demo

Explore the captured Stage 3 results.

Scenario

Loading results

Loading

Benchmark

Solver Comparison

Captured run
Solver Stage Status Cost Penalized Violations Samples

ORBIT

Raw to Repaired

Runtime