Agentic Research Patterns

The 10:40–11:10 block. One simple running example, seven patterns, all live — the whole toolbox in half an hour, right before we put it to work driving three agents in parallel.

A coding agent out of the box is a very fast junior collaborator with no memory, no taste, and no idea what your research standards are. The difference between "neat demo" and "research instrument" is a handful of recurring patterns — ways of structuring the work so the agent's speed compounds instead of dissipating.

We demonstrate each pattern on a single running example, so you see them in context rather than as a feature list.

The seven patterns

Pattern	One-line version	Page
Where agents run	tmux/cluster/remote — decouple the agent's lifetime from your laptop lid	→
Modes	Plan before acting; auto-accept when the loop is safe; spend effort where it pays	→
Subagents	Fan out independent work; keep the main context clean	→
Context & compaction	Context is a finite resource — budget it, compact it, resume from it	→
Skills, MCP, plugins	Teach the agent your lab's conventions once, reuse forever	→
Feedback loops	Give the agent a goal and a measurable loop — this is the big one	→
Evaluation & artifacts	Logs, metrics, plots, reports: make the agent show its work	→

The one-sentence summary

State the goal, give the agent a way to measure progress toward it, and manage its context like the scarce resource it is.

If you internalize nothing else this morning, take that.