One agent or many: how do you split the work?

Blast-radius routing is correct but incomplete — the refinement that changed things for us: declared file manifests. Each agent's task template requires it to list, in its plan, every path it expects to touch. A dumb script diffs manifests across active agents; overlap blocks launch. Sounds bureaucratic, catches a collision a week, and — unexpected bonus — the manifest itself improves the agent's focus. Agents that declare their files wander less. Measurement or placebo, I'll take it.

On agent-to-agent handoff, I've run the experiment properly: agent A implements, writes a handoff doc to a shared file, agent B tests against the doc. Verdict after ~30 tasks: it works only when the handoff artifact is structured (endpoints, types, invariants — a contract) and fails when it's prose. Prose handoffs telephone-game within one hop; A's hedge becomes B's fact. Contract handoffs held up. The human middleman isn't re-explaining, it turns out — they're schema-enforcing. Automate the schema and you can leave.

SRE footnote on the 'practical ceiling': it's not just review capacity, it's incident correlation. When something breaks in staging, one agent's changes are greppable; four agents' interleaved merges took our mean time-to-attribution from minutes to an hour. We now tag every commit with the agent session ID, which helped, but consider incident debugging cost in your ceiling math — it bites exactly when you're least equipped to pay it.

Data point from the other end: we went from four agents back down to two, deliberately. The 2.2x throughput at 4x coordination number matched our experience almost exactly, and two agents at ~1.8x with near-zero coordination overhead wins on any accounting that values founder attention. Parallelism has diminishing returns and attention is the denominator nobody prices.