"More AI agents running doesn't mean there's more of you available - your cognitive bandwidth doesn't parallelize" Watch the full video: https://lnkd.in/gW8nwJHk from our fireside chat at Google I/O We are in a fascinating era of software development where the floor for building has been completely raised. The ability for anyone to spin up many background agents means code can be written while your laptop is closed. But as our toolsets expand, we are slamming into a very human bottleneck: the orchestration tax. Running 20 agents at once might make you feel incredibly busy, but feeling busy is not the same as being productive. When we offload problem-solving without intentionality, we risk falling into cognitive surrender - blindly merging an AI's output without understanding the underlying mechanics. When we stop thinking critically, we lose the ability to debug the very systems we are building. As the role of a developer shifts from writing syntax to writing intentions, we have to evolve our habits to protect our cognitive bandwidth. I personally try to be very intentional with what isolated tasks I delegate to background/cloud agents and which ones I pay more close attention to. How about you? #ai #programming #softwareengineering
The teams that master intentional delegation and protect their cognitive bandwidth today will be the ones consistently building robust systems in 2 years. Others might find themselves constantly debugging agent output they don't fully understand. Addy Osmani
Your insight that feeling busy differs entirely from being productive speaks directly to our need for intentionality, Addy. True emotional safety in engineering cultures develops when teams value deep comprehension over the illusion of parallel progress.
The word fading faster as agentic development accelerates: Accountability.
First of all, it was great to see you on the I/O podium. Congrats! I can’t help noticing a pattern in how we talk about this stuff -and- it doesn’t quite add up for me. AI labs aren’t going to win B2B or B2C adoption with phrases like “cognitive bandwidth” or “cognitive surrender”. None of that tells anyone anything! It’s the same problem across the whole AI industry right now: too much philosophy, too many big words, and not enough clarity. Even as a CTO I find myself asking: what are we actually talking about here. And honestly, if the point is simply that running 10, 20, or 50 AI agents doesn’t automatically make you more effective, then just say that! It’s a good point on its own... It doesn’t need metaphors to land.
Spot on, Addy Osmani The orchestration tax seems like a real problem, but I’d personally argue it’s less about our biological limits and more about interface design . We actually do have the capacity to oversee 20+ agents (MIT’s HAL studies) but in this case the missing link is observability and visualization. I’m convinced the solution lies in what I think of as "flight deck architecture" paired with simple cognitive triggering. If a dashboard can translate complex, multi-agent workflows into simple, binary (green/red) statuses, the human bottleneck disappears. Human-factors research on drone swarm management (like MIT’s HAL studies) proves that when an interface shifts to supervisory control, a single human can comfortably oversee dozens of autonomous assets without cognitive fatigue. We don't need to parallelize our deep thinking; we just need a control tower that lets us make fast, high-level decisions without frying our brains. It’s the exact same principle used in modern nuclear control rooms. Operators don't micromanage thousands of individual valves and pumps; they rely on 'Dark Board' architecture. Everything stays green and quiet until a binary trigger requires human intervention.
Addy Osmani I think being intentional about the specifications, system design, trade-offs is extremely important to avoid compounding problems and cognitive overload at the implementation, review & QA layer. Bringing AI assistance to the system design and planning layer, in a collaborative setting can help offset these problems, since teams could conciously review and approve plans together before changes start. The code output rate might be slower, but I expect the delivery to get faster and quality better.
Great question — and one I think about a lot. I try to be intentional too, but I’ve realized intentionality at the task level is only part of the equation. The bigger challenge is making sure we’re solving the right problems in the first place. Spinning up 20 agents without a clear outcome is like scheduling a meeting with 20 people and no agenda. Everyone’s busy. Nothing gets resolved. Right now there’s enormous pressure to use AI to cut costs, hit KPIs, and show results. But KPIs can be gamed. Outcomes can’t. And chasing the wrong metric with a more powerful tool just produces better-looking garbage. We’ve always known garbage in, garbage out. AI just lets us produce it faster and at scale. What I’ve found is that you have to experience the technology before you can identify the right problems to solve with it. Most organizations and leaders are still in that phase — and that’s okay. So while I agree with being intentional at the task level, I’d push it up a level: how are you assessing which problems are actually worth solving with AI in the first place? That’s where I think the real leverage is — and honestly, where most of us are still figuring it out.
The coordination problem becomes architectural once agents start operating in parallel Addy Osmani. The limiting factor is no longer code generation speed. It’s maintaining intent consistency across dozens of autonomous execution paths. One subtle shift I think we’re underestimating: the developer role is moving from writing implementations to defining invariants. The hard part is no longer getting code produced. It’s ensuring 20 agents don’t each make locally reasonable decisions that collectively create architectural drift. That’s where governance, verification contracts and deterministic constraint enforcement start becoming core infrastructure rather than optional review layers. Running more agents increases output. It doesn’t automatically increase coherence.
Running 20 agents at once doesn't make you a super-developer; it makes you a bottlenecked project manager. The floor for building has been completely raised, but our cognitive bandwidth remains strictly single-threaded. If we offload the thinking process entirely, we lose the exact analytical skills needed to guide the system in the first place. Great reminder that feeling busy with AI automation is not the same as being productive. Intentional delegation is the only way forward.