The Hypothesis

What if an agent could spawn sub-agents that mirror real-world staff and forecast their next appointments?

The Concept

Instead of one central scheduling agent, what if it could boot sub-agents that each mirror a real carer or team on the road? Each shadow agent holds that person's context — their current location, skills, fatigue level, client history — and uses previous data to forecast their next three most likely appointments. The central agent then coordinates across all shadows to build the optimal roster.

Shadow agents

The hypothesis

What if an agent could spawn sub-agents that mirror real-world staff and forecast their next appointments?

The concept

Instead of one central scheduling agent, what if it could boot sub-agents that each mirror a real carer or team on the road? Each shadow agent holds that person’s context — their current location, skills, fatigue level, client history — and uses previous data to forecast their next three most likely appointments. The central agent then coordinates across all shadows to build the optimal roster.

How it works

1. Central scheduling agent
2. Boot shadow agents — one per carer or team
3. Load context per shadow — location, skills, history, fatigue
4. Each shadow forecasts next 3 appointments — based on patterns from historical data
5. Central agent coordinates — resolve conflicts, optimise globally

A fleet of shadow agents that think like individual carers — coordinated by a central brain that thinks about the whole roster.

What it explores

• Can shadow agents accurately forecast appointments using historical scheduling patterns?
• Does a distributed shadow model produce better rosters than a single centralised agent?
• How much historical data does each shadow need before its forecasts are useful?
• Can shadows detect when a carer is likely to burn out based on shift patterns?
• What happens when shadow forecasts conflict — how does the central agent arbitrate?

What we found

• Shadow agents with 3+ months of history forecast the next appointment correctly ~70% of the time
• The distributed model handles last-minute changes better — only affected shadows need to recalculate
• Carers with back-to-back heavy shifts show measurable drops in client satisfaction
• Central coordination is the bottleneck — reconciling conflicts across 50+ shadows is where complexity lives

Learnings

• Distribute context to local agents, coordinate globally — the shadow pattern’s resilience comes from isolating blast radius, not from any single agent being smarter.

• Fatigue is a schedulable variable, not just a human concern — build it into the optimisation function alongside travel time and skill match.

• Conflict resolution across many agents is where decision tree collapse re-emerges — the central coordinator needs a different architecture than the individual shadows.

• The shadow agent pattern generalises to any domain with individual actors and central coordination — logistics, field service, distributed teams.

Where it goes next

This directly informs Wayfinder’s multi-agent architecture. The shadow agent pattern is also relevant beyond rostering — any domain where individual actors have local context that a central system needs to coordinate (logistics, field service, distributed teams) could benefit from this approach.