An agent needs context.
Another needs a routine approval.
A third finishes and waits for its next assignment.
Put enough agents to work and your next bottleneck can be you.
I use Grokbot to manage multiple Codex and Claude sessions via SSH, including by voice from my phone. I did 3 PRs on a 19 mile night run last week in the Utah mountains. It keeps agents moving through those handoffs. Coordination that depended on my attention became part of the system.
That is changing what we should all hire for.
The Multiplier Belongs to the System
We celebrated the mythical 10x engineer. Now I’m starting to think about the 1Mx engineer. I mean an ambition for compounding capability, not a measured millionfold productivity gain.
Exceptional engineers have always made others more effective. The opportunity now is to build systems that carry more of that capability into execution without requiring someone at every step.
A fix becomes a test. A successful approach becomes shared tooling. A coordination failure becomes a workflow that prevents the same mistake across projects.
Two engineers can ship the same feature. One leaves a working feature. The other also leaves tooling that makes future features cheaper to build and easier to verify. Their output looks similar today. Their value separates on the next project, the effects of compounding.
An assessment that only scores the finished feature misses that difference. We start to care less about the product that was built, but the system that built it and the time/effort.
We should value the capability a person leaves behind, not just the work they personally finish. A great tool keeps helping after its author moves to another problem. Some of that tooling improves how the next tool gets built.
I expect the gap between average and exceptional AI developers to widen, the bottom 90% will not be competitive with the top 10%. The same models will not produce the same results when people differ in what they build around them.
Build the Layers That Are Still Human
The emerging stack is being built out of work people still do manually: task assignment, context management, review, coordination, and deciding what happens next.
Every repeated handoff is a candidate for engineering. Managing more agents is an intermediate step. The goal is a system that needs less routine management.
Eventually, I expect a singular coordinating agent: the Joint Master Control Panel. Think J.A.R.V.I.S. from Iron Man. Many agents operate underneath. You direct the system through one interface.
The human contribution shifts toward choosing valuable problems, defining success, and recognizing when convincing evidence is wrong. A system can execute perfectly against a useless objective. Knowing what deserves to be measured becomes more important as execution gets cheaper.
Automating routine approval must not become permission to delete a production database. Boundaries still have to be designed and enforced.
The tasks you are hiring someone to coordinate today may become infrastructure tomorrow. Hire the person who can build the next layer.
Efficiency Is king
Hundreds of agents can produce duplicate implementations, conflicting changes, abandoned branches, and a spectacular bill.
A billion tokens of spaghetti thrown against the wall is not genius.
I care what shipped, whether it works, and what it cost. Measure accepted, tested, useful output against total cost and elapsed time. Include the human effort spent rescuing the workflow.
Then ask what the next project inherits. If every assignment requires the same supervision and repeats the same mistakes, you have increased activity without building much capability.
Genius is becoming a process. The advantage belongs to people who can make that process reliable, economical, and reusable.
Test What the Human Adds
I have spent years building AI systems and working on talent problems. The hardest part is deciding what evidence predicts success months after the interview. Pasted markdown
I call the assessment approach agentic execution benchmarking.
Give candidates realistic problems, modern tools, and explicit constraints. Run a strong agent workflow with comparable tools and resources as the baseline. Matching that baseline may demonstrate competence. Exceptional talent should show where human judgment produces a better result.
For BenchArena, ask for a competitive new algorithm under 40MB. Define the resource limit precisely. Record performance and the tokens spent discovering it.
For a database role, use a bounded SQL implementation competing with SQLite on a declared workload. Test correctness and crash recovery alongside speed. A benchmark win is worthless if the database loses your data.
Ask candidates to demonstrate a harness innovation: better tooling for assigning, coordinating, or verifying agent work. Require evidence that it improves the workflow.
Then give them an unseen second problem.
The first tests execution. The second helps reveal whether they built reusable capability or an elaborate solution to one demonstration. Compare against the baseline again. Did quality improve or total cost fall without adding hidden supervision? Change a requirement or break a dependency. Watch how they diagnose the failure and adapt.
Use bounded, paid work samples with comparable tools and budgets. Candidates can propose evaluation criteria. They do not get to mark their own homework.
Keep the agent baseline current. Validate the assessment against performance six to twelve months later. Audit whether it disadvantages candidates for reasons unrelated to the work.
An impressive sandbox is not proof of a useful hiring signal.
Let the Evidence Surprise You
I honestly do not know whether the neuroscience PhD or the Uber driver I met in Austin would be the better hire for this work. I want an assessment capable of finding out.
Drive matters. Obsession matters. Look for obsession with improving the system, not merely spending more hours inside it.
Keep the resume for context. Stop treating it as the verdict.
The 1Mx engineer is not defined by how many agents they command. Look at the capability that accumulates, the people it helps, and the work it makes possible.
Do not just measure what someone can produce.
Measure what becomes possible because they were there.
