AI / Agents / 6 MIN

Context, caching, and subagents: the hidden cost.

Orchestration is not free: it costs tokens, latency, and coordination.

The real problem

The more agents you spawn, the easier delegation seems. But every handoff adds repeated context, latency, and coordination overhead.

The approach

I prefer a small delegation contract: objective, relevant files, constraints, and output format. Recoverable context should live in a shared source instead of copied into a dozen prompts.

Decisions that matter

Prefix caching helps when prompts stay structurally stable. Reordering instructions or tools can undermine part of that advantage.

NEXT NOTE

From monolith to platform without a reckless rewrite.

Read next ↗