mittuled manager-development-programme
Agent: Agent Trainer & Skill Optimizer
L2 Agent Trainer and Skill Optimizer (1x) responsible for evaluating existing skills, running evals, optimising prompts, improving agent performance, and training new agent configurations.
Department ethos: ideal-agent-operations.md
Skill Description
Trains orchestration agents on delegation, context management, and multi-agent coordination to improve fleet-wide workflow execution.
When to Use
- When orchestration agents are making poor delegation decisions, routing tasks to wrong agents or overloading specific agents.
- When new agents have been added to the fleet and orchestrators need updated awareness of available capabilities.
- When multi-agent workflows are failing due to context loss during handoffs between agents.
Workflow
- Assess Orchestrator Performance: Evaluate current orchestration agents on delegation accuracy, context preservation during handoffs, and coordination efficiency. Deliverable: orchestrator performance baseline.
- Identify Training Gaps: Analyze delegation errors, context loss incidents, and coordination failures to determine specific training needs. Deliverable: training gap analysis.
- Design Training Programme: Create training scenarios covering delegation logic (which agent for which task), context packaging (what to include in handoffs), and coordination patterns (sequential, parallel, fan-out). Deliverable: training programme with scenario sets.
- Execute Training: Run orchestration agents through training scenarios. Evaluate outputs against expected delegation decisions and context handling. Iterate on prompt updates and configuration changes. Deliverable: training results with before/after comparison.
- Validate and Deploy: Confirm improved orchestration performance meets thresholds. Deploy updated configurations to production orchestrators. Deliverable: validated orchestrator configurations.
Anti-Patterns
- Training on happy paths only: Practicing delegation with clear-cut scenarios but never with ambiguous or multi-valid-answer tasks. Why: production orchestration frequently involves ambiguous routing where multiple agents could handle the task; training must cover these cases.
- Ignoring context window budgets: Training orchestrators to pass full context without considering downstream agents' context window limits. Why: oversized context causes truncation or overflow, leading to lost information at the worst possible moment.
- Static training: Training once and never refreshing as the fleet evolves. Why: the orchestrator's knowledge of available agents and their capabilities becomes stale, leading to misroutes.
Output
On success: Produces updated orchestrator configurations with improved delegation logic, context management, and coordination patterns. Delivered as deployed configurations with performance improvement metrics.
On failure: Report which training scenarios the orchestrator still fails, what prompt or configuration changes were attempted, and what additional training data or scenarios are needed.
Related Skills
onboarding-programme-builder-- New agent onboarding feeds into orchestrator awareness updates.skills-gap-analyser-- Skill gaps in orchestrators are a specific case of the broader gap analysis.team-health-monitor-- Coordination failures surface as health monitoring anomalies.