2026-05-10
May 10, 2026
The "manager-as-AI-orchestrator" framing dominating Q2 keynotes mistakes the surface for the substance. The real leverage moves to leaders who can encode judgment in durable text — strategy memos, working-backwards docs, decision records — that humans and agents can both reason against.
There is a story being told at every engineering-leadership conference this quarter. It goes like this: the manager of the future is an "AI orchestrator," conducting a symphony of agents and a smaller, more senior human team. Microsoft's 2025 Work Trend Index is the sermon text — 82% of executives expect AI agents in their workforce within 18 months, only 23% are confident they can integrate them well. The implication is that the gap between those two numbers is a tooling problem.
It isn't. The gap is a judgment problem, and judgment in 2026 lives in writing.
The "orchestrator" framing is a category error. Orchestration is a technical implementation skill — pipelines, MCP servers, agent graphs, eval harnesses. It's important, and your platform team should own it. But it is not the seat of technical leadership any more than running a CI pipeline was the seat of leadership in 2010. The thing that determines whether your AI-augmented org compounds value or burns headcount is upstream of the orchestrator: it is the quality of the written artifacts that encode what the org is for and how it decides.
Three forces are converging this year, and most "future of leadership" essays are reading two of them and missing the third.
Force one: mid-management is being compressed. "Manager of managers" is the single line item under sharpest scrutiny in 2026 cost reviews. The narrative that a quarter of recent tech layoffs are explicitly AI-attributed is doing real work in budget meetings. Director scope is widening; the org is flattening from below.
Force two: AI agents are eating the routine coordination work. Sprint coordination, status synthesis, dependency-graph upkeep, RAID-log hygiene, meeting summarization, action-item tracking — all of it is being absorbed by tools that didn't exist 18 months ago. This is not a prediction. Senior TPMs reading this can already point to the four ceremonies they no longer run themselves.
Force three — the one most essays miss — is that the artifacts agents need to do that work do not yet exist in most organizations. An agent cannot run your standup if your team has no written charter. An agent cannot triage your backlog if your prioritization criteria live in three Slack threads and one Director's head. An agent cannot draft your weekly exec brief if your program goals are encoded only in a deck nobody has opened in six weeks. The ceiling on AI leverage in your org is the floor of your written substrate.
This is why the "orchestrator" job description is wrong for senior leadership. The orchestrator role assumes the substrate exists. In most large eng orgs in 2026, it doesn't — and building that substrate is the most defensible thing a senior TPM or tech leader can do this year.
I'll call the resulting archetype the writer-executive, with apologies to Will Larson, who would rightly point out it's a special case of the Right Hand. The writer-executive's daily output looks like this:
None of this is new. What's new is that these artifacts have stopped being optional. They are the API your AI tools — and your hybrid team — read against. Writing them is no longer "documentation work"; it is the substantive output of senior leadership.
There is a hierarchy of what AI can replace, and it tracks the hierarchy of what is and isn't written down. Verbal coordination — the stuff that lived in your head and your hallway conversations — gets eaten first. Stale written artifacts get refactored next. Living, versioned, decision-grade documents are the last thing to be touched, because they are the source of truth the tools work from. The leaders who survive the next two years are the ones whose calendars look unrecognizable to a 2023 director: less time in meetings, more time writing, and far more time reviewing decision documents drafted by their team and their agents.
The reason most engineering leaders don't make this shift is that writing is uncomfortable. Writing exposes sloppy thinking. A 60-minute meeting can be 60 minutes of fog. A six-page narrative cannot. Anyone can become an orchestrator with six months of practice on the right stack. Far fewer people can write a strategy memo a CTO will sign. The orchestrator skill set commodifies on the same curve as the tools — fast. The writer-executive skill set compounds.
The forecast for the next decade of technical leadership, then, is not "managers become orchestrators." It is: the value of an engineering leader is increasingly equal to the quality of the written artifacts they own. The IC engineer's edge migrates from typing speed to taste. The leader's edge migrates from meeting cadence to memo cadence.
Your AI tools will get better. Your org's appetite for AI leverage will grow. The question for the next 18 months is whether your written substrate grows fast enough to catch it. If it doesn't, your team will spend 2027 retrofitting clarity onto decisions made in haste in 2026 — and the writer-executives at the company you almost joined will be three quarters ahead.
Try this week. Pick the single most important program you own. In 90 minutes, no more, write a one-page charter: Problem, Customer, Non-Goals, Success Criteria with thresholds, Single Approver, Dated Last-Reviewed. Don't socialize it yet. Read it back in 24 hours. If you can't honestly say it would let an outsider operate the program for a week without paging you, the program isn't ready to be augmented by anything — agent or human. Fix the document, not the team.
What it is. Amazon's pre-build forcing function, formalized inside the company in the early 2000s by Colin Bryar and Bill Carr and described publicly in Working Backwards (2021). Before any meaningful build investment, the team writes a one-page mock press release announcing the finished product, plus an FAQ that pre-empts the obvious internal and customer objections. The PR/FAQ is read in silence at the start of the kickoff meeting; only then is it discussed.
When to use it. Any initiative where the cost of building the wrong thing is higher than the cost of writing the right doc. New platform investments. Re-architectures whose value depends on a behavioral change you haven't tested. AI-feature launches where the customer-facing claim is unproven. Cross-team programs where the prize is integration, not code.
How to run it:
When NOT to use it. Sustaining engineering work. Bug fixes. Anything reversible inside a single sprint. The discipline is expensive; reserve it for decisions you can't unwind.
One-line example: the team that built AWS Lambda is widely reported to have spent more time on the PR/FAQ than on the prototype. The doc, not the code, was the thing that survived contact with internal review.
How agentic AI will reshape engineering workflows in 2026 — CIO's framing matches the orchestrator narrative critiqued above, but the data on the role transition is honest: engineers shifting from authoring code to designing system architecture, defining objectives and guardrails, and validating output. Read it as a diagnostic of where your team is, not a prescription.
5 uncomfortable predictions for engineering leaders in 2026 — LeadDev's annual list. The most pointed prediction: the "Senior-Only" hiring pattern produces an inverted pyramid that breaks legacy systems within 24 months. If your org has frozen entry-level hiring, this is the risk to surface in your next planning review.
Building internal agents — Will Larson's running series on shipping agents inside engineering orgs. Worth the subscribe even if you delegate the building. The framing — agents as scaffolding for recurring work, not as autonomous staff — is the most useful corrective to the orchestrator hype.
"There is no way to write a six-page narratively structured memo and not have clear thinking."
— Jeff Bezos, 2017 Letter to Shareholders
Don't miss what's next. Subscribe to Critical Path: