Organisational design · Research platform

SuperScript — reinforcement learning for team allocation

A multi-year research platform pairing an agent-based model of team formation in project-based organisations with a suite of team-allocation optimisers — including hierarchical reinforcement learning — benchmarked head to head across organisational scenarios.

Four radar charts comparing the best team-allocation solutions found by different optimisers — Greedy, GRASP, hierarchical RL and an ensemble — on six axes (overall value, risk, team fit, individual fit, creativity, skill balance) across four scenarios; within each scenario the optimisers' shapes almost completely overlap. — Best solutions per optimiser, averaged over the test dataset, across four scenarios (linear / nonlinear success function; 100 / 20 workers). Short-term success probabilities are near-identical between methods — the differences emerge in the organisations they build.

The problem

Allocating people to teams in a project-based organisation looks like a scheduling puzzle but behaves like an ecosystem. Skills, availability, project requirements and team chemistry all interact — and every allocation reshapes the organisation that receives the next one. Most allocation research stops at the first question: which assignment scores best right now? SuperScript was built to ask the harder second one: what kind of organisation does each allocation method build over time?

The platform

At the core is SuperScript itself: an agent-based model of team formation in project-based organisations, built with Mesa. Simulated workers carry skill profiles; projects arrive continuously; teams are assembled, deliver or fail, and the workforce evolves in response.

On top of the model sits a suite of six team-allocation optimisers, benchmarked head to head: Random and Greedy baselines, GRASP, mixed-integer linear programming (solved with SCIP), hierarchical reinforcement learning — a higher-level worker-and-team-selection policy (RLD2) sitting over a lower-level hard-skill allocator (RLD1), trained with Gymnasium and Stable-Baselines3 PPO — and an Ensemble combining methods.

What the experiments show

Run across scenarios — linear and nonlinear project-success functions, organisations of 20 and 100 workers — the optimisers reach strikingly similar short-term success scores. The interesting differences are structural: methods with near-identical scores assemble very different teams, and over time produce very different organisations. Team allocation, in other words, is organisational design.

The platform is open: three public repositories carry the agent-based model with its optimisers, the RL environments with trained policies, and the full experiment suite — including a golden regression suite that verifies the results reproduce. A journal article is in preparation.