Buy owner bandwidth before agent seats.

Video caption

Buy owner bandwidth before agent seats.

Output counts hide waiting and rescue.

Routine, exception and novel design need different gates.

My rule: Agent concurrency follows an observed release gain.

#EricFieldNotes

Full written post / accessibility read

If agents now produce the bulk of patches, the scarce capacity shifts toward product, architecture and quality judgment. Adding sessions without changing that decision system can turn a team into a warehouse of plausible unfinished work.

Management sees pull requests opened, tests run and tokens consumed. It may not see how long a product owner waited to resolve a tradeoff, or how many changes had to be rebuilt after a late decision. Those hidden intervals govern actual release speed.

Write routine architecture and quality constraints into a trusted harness. Route true policy exceptions to a named owner with a compact decision packet. Reserve deep design review for novel, high-impact changes. Measure accepted changes per week, queue age, review load and escaped defects by lane.

Pilot more agents only where the acceptance lane can absorb them. If queue age or rework rises, improve policy clarity and review flow before adding sessions. Do this because the business receives accepted outcomes, not a high count of generated patches.

#EricFieldNotes

Four-beat scene transcript

1. Buy owner bandwidth before agent seats.

If agents now produce the bulk of patches, the scarce capacity shifts toward product, architecture and quality judgment. Adding sessions without changing that decision system can turn a team into a warehouse of plausible unfinished work.

Visual: A faster code factory can starve the decision gate.

2. The familiar dashboard rewards the wrong thing.

Management sees pull requests opened, tests run and tokens consumed. It may not see how long a product owner waited to resolve a tradeoff, or how many changes had to be rebuilt after a late decision. Those hidden intervals govern actual release speed.

Visual: Output counts hide waiting and rescue.

3. Give every task an acceptance lane.

Write routine architecture and quality constraints into a trusted harness. Route true policy exceptions to a named owner with a compact decision packet. Reserve deep design review for novel, high-impact changes. Measure accepted changes per week, queue age, review load and escaped defects by lane.

Visual: Routine, exception and novel design need different gates.

4. Scale accepted decisions.

Pilot more agents only where the acceptance lane can absorb them. If queue age or rework rises, improve policy clarity and review flow before adding sessions. Do this because the business receives accepted outcomes, not a high count of generated patches.

Visual: Agent concurrency follows an observed release gain.

Research and claim limits

• METR randomized trial (S20)
• METR 2026 productivity experiment update (S93)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.