Product Trio in the Agentic Era

How the PM / Designer / Engineer trio is evolving as AI agents absorb execution work — and what each role actually does now.

Last updated: 2026-04-24

Living document. We are in active transition. This page will be revised as patterns solidify. Claims marked with ⚠️ are contested or unresolved.

What Changed and What Didn’t

The classic product trio (PM, Designer, Engineer) was designed around a monthly software cycle: PM writes spec → Designer mocks → Engineer builds → repeat. AI collapses that loop. Prototypes take hours, not sprints. A solo engineer with agents can ship in a day what previously required a coordinated team.

What changed:

Execution cost dropped by 10–100x — the bottleneck is now judgment, not labor
Blurred functional identity — engineers design, designers write copy, PMs analyze data and ship prototypes (Zhuo: “identify as a problem-solver, not a role”)
Discovery acceleration — agents can synthesize research, draft interview guides, analyze patterns across hundreds of conversations
Spec format shifted — a well-written CLAUDE.md or skill file is now the most important product artifact, not a PRD

What didn’t change:

Someone must talk to customers — agents don’t replace customer contact; they can help synthesize what you hear, not generate what you need to hear
Judgment about what to build — commercial instincts, prioritization, knowing which problem matters most — still human
Organizational navigation — in big companies, half the job is alignment, trust, and politics; agents don’t help here
Accountability — someone still owns the outcome

Context A: Small Team Inside a Big Company

Assume: 3–8 people, part of a larger organization, some autonomy but real stakeholder relationships, existing systems and customers

PM

What shifts: Less time on documentation and synthesis (agents do first drafts of specs, research summaries, competitive analysis). More time on three things agents cannot do:

Customer contact — being the person in the room (or call) who is genuinely listening, not validating. Torres’s research is clear: teams that sustained discovery throughout succeeded; domain expertise helps but doesn’t replace contact.
Organizational navigation — securing mandate, managing stakeholders who give input vs. dictate, running the pilot-team playbook (Cagan) in an environment that defaults to project mode
Commercial judgment — knowing which customer problem is worth solving, what outcome to chase, whether the business model holds

What accelerates: Prototype-and-prune becomes viable at smaller team size. A PM who can write clear briefs and evaluate outputs can run 3–5 prototype experiments in the time it used to take to write a single PRD.

The new failure mode: Mistaking output volume for discovery quality. Agents can generate insights that sound real but aren’t grounded in actual customer contact. Usage data ≠ customer understanding.

New skill worth building: Writing prompts and specs that produce useful agent outputs. Not engineering, but something adjacent — clarity of thought made machine-readable.

Designer

What shifts: Less time on execution (generating variations, resizing assets, prototyping flows). More time on:

Taste — distinguishing good from excellent when you can generate 50 variations in an hour. This is the scarce resource. Field’s framework applies: taste (filtering preferences), craft (knowing which micro-decisions matter), point of view (what makes this ours).
Evaluation — the design review loop is now high-frequency. Agent-generated designs need a trained eye at every step, not just at the end.
Defining the diverge/converge contract — when to expand possibility space (diverge) and when to commit (converge). Agents are great at generating; humans still need to call the prune.

What accelerates: Figma Make data suggests 60% of designs already come from non-designers — the designer’s job shifts from sole maker to quality gatekeeper across a wider surface area.

The new failure mode: Losing the ability to make things by hand. Taste requires a making practice. If the designer only evaluates, they drift from the craft that makes evaluation trustworthy.

Practitioner evidence (Felix Lee, ADPList CEO): Built three production apps over 3.5 months entirely in Claude Code — never opened Figma. His conclusion: Figma is becoming a collaboration/design system layer, not a creation surface. The designers who will matter are those who develop Claude Code fluency, not those who stay Figma-only. Most designers today are “not freaked out enough” — still on traditional workflows, adoption of deep AI tools up only 10–20% in 6 months.

⚠️ Open question: Does the designer role consolidate with research (since both are now judgment-heavy pattern-recognition jobs)? Or do they stay separate?

Engineer

What shifts: Less time on typed-out implementation, more time on:

Architecture and constraints — defining what agents can and can’t touch. The harness (Agentic Engineering’s Configuration + Guardrails layers) is now a core eng deliverable.
Evaluation — running evals on agent output. Agents can write code; engineers judge whether it’s correct, safe, and production-grade.
Skill and tool design — writing skill files, MCP tools, and agent configs is now a primary engineering output. The meta-rule from OpenClaw applies: if something will happen again, codify it.

What accelerates: A single engineer can now maintain systems that previously required a team. The surface area of what one person can own expands significantly.

The new failure mode: Shipping agent-generated code without adequate eval coverage. Agents are fast and confidently wrong.

The Trio as a Unit

The trio’s greatest shift: from serial handoff to parallel experiment. Instead of PM → Design → Eng, the pattern is now:

Everyone can prototype
Short daily loops on what works
Prune relentlessly; ship survivors

The coordination overhead drops, but taste arbitration rises. Someone has to have the final call on “is this good enough?” — in practice, this is usually the PM or designer, but it needs to be explicit.

In big-company context, the trio also needs a shared answer to: how do we show our work to stakeholders without showing every prototype iteration? Cagan’s answer: run discovery demos that show the thinking, not the transcripts.

Context B: Company of < 10 People

Assume: early-stage or small startup, everyone wears multiple hats, no separate discovery function, probably a founder or two

The Trio Collapses

There likely isn’t a dedicated PM, Designer, and Engineer. There might be 1–2 engineers who also do product decisions, a founder who also does customer development, and a designer if you’re lucky.

This is fine. The trio was always a minimum viable team for sustained product work, not a fundamental law. What the roles represent still matters:

What the role represents	Who owns it in a < 10 person team
Customer contact + judgment about what to build	Founder / CEO, or a dedicated “customer-obsessed” person
Taste + visual and UX quality	Designer if you have one; otherwise the person with the strongest opinions about “good”
Technical feasibility + system integrity	Lead engineer or whoever owns the codebase

The risk is that all three collapse onto one founder who is also shipping code, and customer contact becomes the thing that falls off first because it’s the hardest to schedule and the least urgent in the short term. Torres’s research confirms this: it falls off even when teams know they should be doing it.

Every Person is a Barrel

At this scale, Rabois’s barrels-and-ammunition framing is existential. Every person must be a barrel — able to take an outcome and deliver it without hand-holding. Hiring ammunition (even excellent ammunition) before you have enough barrels is fatal. One barrel can direct multiple agents; one ammunition hire without a barrel to direct them is drag.

This reframes the hiring question: not “do we need a PM or designer?” but “does this person own outcomes end-to-end, or do they need direction?”

What Agents Actually Give You at This Scale

Execution leverage: one engineer with agents can ship at the pace of a 3–5 person team. This is real.
Generalist enablement: a founder who isn’t a designer can generate credible design explorations and have a designer evaluate them, vs. waiting 2 weeks for design bandwidth
Documentation and synthesis: agents write specs, summarize research, maintain wikis (this wiki is that pattern)
Parallel exploration: prototype-and-prune is actually more tractable for a tiny team than for a 20-person team because there’s less coordination overhead per experiment

What Agents Don’t Give You

The founder’s customer conversations — this is still the most valuable input and the hardest to automate. Agents can analyze call transcripts; they cannot build the relationship that produces honest answers.
Commercial judgment — knowing which bet to make when you have 5 reasonable options. This is pattern recognition from deep domain + market experience, not prompt execution.
Taste development — if nobody on the team has taste, agents produce mediocre work faster. Quality of output is ceiling’d by quality of judgment.

Practical Role Sketch for a 3–5 Person Team

Person 1 (Founder/CEO): owns customer contact, what-to-build judgment, and the external narrative. Directs agents for research synthesis, competitive analysis, and first-draft specs. Does not code, unless technically founding.

Person 2 (Lead Engineer): owns the technical architecture, the harness, and evals. Uses agents heavily for implementation. Owns skill files and agent configuration. Reviews all agent-generated code before ship.

Person 3 (Designer or PM): owns taste and the quality bar for shipped output. If designer: evaluates agent-generated UIs, runs the diverge/converge cycle, develops and protects the visual and UX POV. If PM: runs discovery loops, writes briefs that agents can execute from, manages the prototype-and-prune backlog.

Agents (treated as junior generalists): write first drafts, generate variations, do research synthesis, maintain documentation, run cron-based monitoring, draft communications for human review.

Roles Under Pressure (Both Contexts)

Role	Pressure	Why
Information-mover PM	Critical	The whole job automated — status reporting, synthesis, roadmap framing, alignment theater
Junior PM	High	First-draft synthesis, ticket writing, research summaries — agents do these faster
QA Engineer	High	Agents write and run tests; humans review coverage strategy
Pixel-generator Designer	High	Variation generation, resizing, production handoff absorbed by agents
Junior Designer	Medium	Production work automated; taste/evaluation capacity still developing
Senior PM (judgment)	Low	Judgment, discovery, organizational navigation remain human
Senior Engineer	Low	Architecture, evals, and agent orchestration require deep experience
Designer with strong taste	Low	Taste is the scarce resource; demand goes up as supply of generated design explodes
Researcher	⚠️ Unclear	Some synthesis work automated; but actual customer contact is irreplaceable

The core pattern (Singhal): the information-mover / coordination layer across all roles is being automated. The judgment / taste / customer-contact layer is becoming more valuable and more explicitly demanded.

Practitioner Evidence: Block’s 40% RIF (Owen Jennings, 2026)

Block (Square/Cash App/Afterpay) executed a ~40% reduction in force in early 2026. Owen Jennings — their business lead who oversees product operations and customer support — is explicit that this was not financial overhang. The cuts were heaviest on the development side; outbound sales and account management were barely touched. This is the opposite pattern of a typical budget-driven RIF.

The trigger was the December 2024 inflection (Opus 4.6 + Codex 5.3): models became capable of working with existing complex codebases overnight, not just greenfield. The correlation between headcount and output that had held for decades “basically broke the first week of December.”

What the new org looks like at Block:

Squads of 1–6 replacing feature teams of ~14
50–60% fewer management layers on the development side; product org has 2–3 max
All designers and PMs shipping PRs (no longer notable)
70–80% meeting reduction — time returned to building
New work pattern: “In the background, 10–20 agents are doing a bunch of stuff, and I have to check in, nudge, change, commit.” Not linear workflow — supervision + editing loop.
Fluid squad assignment: move between products across cycles instead of permanent team membership

The compliance team was not significantly cut — regulatory relationships require human accountability and asymmetric risk tolerance.

“We’re not writing code by hand anymore. That’s over. That’s done.”

See owen-jennings for the full picture.

Practitioner Analysis: Nikhyl Singhal (2026)

Nikhyl Singhal — CPO veteran at Meta/Google/Credit Karma, runs the Skip community of ~125 heads of product — offers the most data-grounded account of what’s actually happening to the PM function. He talks to active product leaders continuously, which makes his framing more concrete than most.

The 50/50 Split: Information Movers vs. Builders

The core diagnosis: about half of current PMs are in serious trouble. The at-risk half are the “information movers” — people whose job was carrying information between organizational layers: reframing team output for the manager, reframing for the manager’s manager, status reports, alignment theater, roadmap decks. Responsibility without authority.

This was always a thin definition of the PM job. AI is now automating it entirely. The ground-truth information that used to require a human conduit (what is the team working on? what does this customer want? what’s the status of this initiative?) can now be surfaced directly, without spin, by anyone who asks.

The surviving half — builders with judgment — are having the time of their lives. Direct connection from idea to shipped product. No more waiting for designer queues, engineering backlogs, or approval chains. The product person who wanted to build but was stuck in a coordinator role can now act on that instinct.

The 30,000 / 8,000 Prediction

12–24 month forecast: large companies shed 30,000 and rehire 8,000. The 30,000 are a combination of ZIRP-era overhiring that never delivered proportionate output plus skills that don’t translate. The 8,000 will all be AI-first builders.

This happens simultaneously with a 3+ year high in open PM roles (as of April 2026). The paradox resolves: demand for builder PMs is up; demand for information-mover PMs is collapsing.

Block’s 40% RIF (Owen Jennings, 2026) is the live example. Cuts heaviest on development side; the correlation between headcount and output “basically broke the first week of December” (the December 2024 model capability inflection). See owen-jennings.

Logo Depreciation

The ZIRP career playbook: accumulate prestigious logos (FAANG), demonstrate scale experience, leverage brand for advancement. This is inverting.

Six years at a large company working in non-AI-native ways is now a liability. You come out unable to describe your work in terms that resonate with where product is heading. The hiring question has shifted from “where did you work?” to “how modern are you?” and “show me your judgment.” Companies are scenario-testing in interviews: what tools do you use? how do you think? — not “walk me through the product you shipped in 2022.”

PM as Dandelion Seeds

Long-run optimism: product builders become the change agents every industry needs. Product leaders are the first function to master AI-native building. Within 2–4 years, marketing, sales, operations, and non-tech industries (healthcare, education, manufacturing) will desperately need people who can bring obsolescence thinking, judgment, and builder skills into their organizations. PMs who master this first become the consultants and transformation leaders the rest of the economy hires. “They’re like dandelion seeds when you blow it and they just go everywhere.”

The Reinvention Psychology

The barrier to transition isn’t skill — it’s psychology. Four compounding factors:

Power years collision: people in their 30s are at peak career competence and peak life obligation (kids, aging parents, health, financial pressure). The 8–12 available hours are already overcommitted.
Shadow superpower: the better you were at mastering the old system, the less incentive to recognize the new one. Employer still sees strong performance. Nothing feels broken.
Moving target: unlike normal transitions (grind for a year, reach the new state, settle), this requires continuous reinvention. Falling behind for 3 months is meaningful; the game keeps changing.
Change aversion: once humans find a working system, they optimize for stability. Kids fall down constantly; adults stop falling on purpose.

The crossover mechanism: joy. Everyone who successfully transitions has a “first joy” story — building a chief-of-staff app, automating a tedious workflow, making something their partner actually uses. That one successful build creates the feedback loop. Joy is the antidote to burnout. Leaders’ job: manufacture moments of joy for their reports.

Design Parallel: Pixel Generators vs. Taste Makers

The information-mover / judgment-PM split maps directly to design. “Pixel generators” — designers whose value was production throughput (variants, resizing, handoffs) — are in the same trouble as information-mover PMs. “Taste makers” — designers who bring aesthetic judgment, strong POV, and evaluation skill — are increasingly valuable.

The risk: most hiring plans conflate the two. Companies automating pixel generation with AI tools but not recognizing that taste is a different and irreplaceable capability. The plateau in design hiring (as of early 2026) may reflect this confusion more than it reflects genuine declining demand for taste.

Practitioner Observation: “Workarounds Are the New Workflows”

From Figma x Anthropic live session (Brett, Figma designer advocate):

“If there’s one thing I’ve learned in the last six, seven, eight months is that workarounds are kind of the new workflows — and I don’t say that disparagingly. Everyone’s figuring out how to stitch this stuff together and get good results.”

This is the honest state of PM/Design/Eng collaboration in early 2026. Nobody has a settled playbook. The people winning are:

Curious and collaborative — learning from adjacent roles, not waiting for a defined process
Horizontal — designers making PRs, engineers caring about design quality by default
Building shared skills — using skill files as cross-functional knowledge transfer artifacts

Tariq (Anthropic): “The best engineers are doing 10% coding with Claude Code and 90% exploration, artifact creation, and the busy work that’s not as useful.” The same is increasingly true for designers and PMs.

Open Questions (As of 2026-04-24)

Who owns the agent stack? — skill files, eval suites, harness configs. Is this PM, Eng, or a new role (“Agent Ops”)?
Does continuous discovery survive prototype-and-prune? — Zhuo dissolves the PM into a generalist; Torres says customer contact is load-bearing. Which model wins, and does context (B2B vs B2C, early vs. late stage) determine it?
Where does taste live? — Is taste a skill that can be learned, or is it a trait you hire for? If everyone can generate, and taste filters generation, does taste become the #1 hiring criterion at every level?
Does the trio become a duo? — In a < 10 person team, Designer + Engineer (both capable of product judgment) may replace the classic three-role structure, with customer contact owned by the founder outside the product trio.
How do you do discovery at agent speed? — If you can prototype in hours, interview cadence is the bottleneck. Can agent-assisted interview synthesis (transcript analysis, pattern extraction) compress the discovery loop without losing the signal that comes from being present?
What does alignment look like once ground truth is easy? — Singhal: a lot of “alignment” was just getting people the same ground-truth information through all its spin layers. AI removes the spin; the CEO can directly ask their agent for ground truth. But someone still has to fight for an opinion and decide. Does easier ground truth make alignment faster or just expose the underlying political reality more nakedly?
Will PM “dandelion seeds” spread to non-tech industries? — Singhal’s prediction: PMs who master AI-native building become the change agents every sector needs. Plausible but unverified. The constraint may be domain knowledge — being great at AI-native building doesn’t automatically give you enough healthcare/education/manufacturing context to lead transformation there.

Connections

product-operating-model — Cagan’s empowered trio framework; still the baseline being disrupted
prototype-and-prune — Zhuo’s replacement model; directly challenges the trio’s serial handoff pattern
thin-harness-fat-skills — engineering’s new primary output is skill files and agent configuration, not just shipped features
agentic-engineering — four-layer framework for what engineers now configure
barrels-and-ammunition — critical lens for small teams: agents amplify barrels, don’t substitute for them
auto-research — the discovery acceleration pattern; partial answer to “who does research?”
design-taste-craft — Field’s taste/craft/POV framework; names what designers protect in the agentic era
agent-first-software — the end-state this team structure is building toward
teresa-torres — continuous discovery; customer contact is load-bearing even when everything else is automated
julie-zhuo — prototype-and-prune; role dissolution; taste as bottleneck
keith-rabois — barrels/ammunition; PM dissolution; undiscovered talent at smaller scale
owen-jennings — Block practitioner; 40% RIF; squads of 1–6; supervision loop; December 2024 inflection
nikhyl-singhal — information-mover PM diagnosis; 30k/8k prediction; logo depreciation; PM dandelion seeds; reinvention psychology
product-development — broader topic area
ai-agents — the agent layer that changes the math

Sources

This is a synthesis page. No single source; draws on all ingested material as of 2026-04-24.

second-brain

Explorer

product-trio-agentic-era