Agentic Patterns

Structural patterns for organizing agentic workloads. These are the building blocks for how work gets distributed, monitored, and completed across agents and sessions.

The core question is always the same: who does the work, at what cost, with what model, and how do you know when it's done? The patterns here are different answers to that question depending on the workload characteristics.

Patterns

• Scout / Dispatch — cheap model scans for work, expensive model only fires when needed
• Specialized Subagents — purpose-built agents with locked-down tools and focused context
• Research Fleet — parallel subagents for broad exploration with executive summary aggregation
• Multi-Stage Pipelines — triage → design → execute with separate models and cadences
• Chained One-Shot Crons — ephemeral cron chains that self-assemble, verify, retry, and self-terminate
• Fire-and-Forget Dispatch — executor dispatches work and exits; state tracked externally
• Orchestration Transparency — mirror all dispatches to a monitoring channel for full visibility

Choosing a Pattern

Pattern	Best for	Hit rate	Cost profile
Scout / Dispatch	High-frequency monitoring with low signal	<40%	Cheap most runs, expensive when needed
Specialized Subagent	Recurring task type needing isolation	N/A	Per-spawn, predictable
Research Fleet	Broad exploration across dimensions	100% (always produces output)	Burst cost, fast wall-clock
Multi-Stage Pipeline	Continuous delivery with quality gates	Varies per stage	Spread across stages
Chained One-Shot	Bursty incidents needing verify/retry/escalate	Event-driven	Temporary burst, auto-decays
Fire-and-Forget	Batch dispatch of independent work items	100%	Executor is cheap; workers are the cost
Orchestration Transparency	Any of the above	N/A	Near-zero (just message mirroring)

These patterns compose. A pipeline's executor stage might use fire-and-forget dispatch, which spawns specialized subagents, with orchestration transparency logging every dispatch.