What is an agentic pod?
An agentic pod is a small, permanent team of people and AI agents that owns one part of a business. It pairs a domain expert who knows the work with a builder who can ship agents, under a named sponsor who is accountable for the result.
Unlike a project team, a pod is permanent: it keeps the workflow it automates, accumulates agents over time, and reports one number for what it costs and one for what it saves. Uber's CTO popularized the term in July 2026. A services firm had been using it publicly since April 2025.
2 days to 10 min
Financial pacing reports after one Uber pod. Uber's CTO
150,000+ agents
Per average Fortune 500 enterprise by 2028, up from less than 15 in 2025. Gartner
More than 2x
Organizational factors beat individual ones for AI impact, 67% vs 32%. Microsoft
Where the term came from
The term has a public history, and it is more interesting than a single coinage. Andres Angelani, CEO of the engineering firm Wizeline, was writing about Agentic Pods in April 2025, anchored to the agile pods his industry already knew, and by June 2025 had defined them as cross-functional units combining human expertise with AI agents. Fifteen months later, the CTO of a public company reached the same words independently.
“So we created something called Agentic Pods. The idea is simple. We handpicked ~30 of our most AI-proficient engineers and paired each of them with a domain expert from a business function.”
Uber ran sixteen pods across sixteen business functions in two months: finance, legal, operations, marketing, customer support, HR, and procurement. Business Insider covered it within 48 hours. Within eight days, Kieran Flanagan, SVP of Agentic GTM at HubSpot, had adopted the model for go-to-market teams and written the cleanest one-line version anyone has published:
“From sprawl to system: agentic pods. The concept is simple: pair an AI-proficient engineer with a domain expert, give them a tight window, and ship a working end-to-end workflow.”
Two parties converging on the same name, fifteen months apart, is better evidence that the name is natural than one party inventing it.
What is inside a pod
A pod is three roles and a roster of agents. The roles are stable; the agents accumulate.
The sponsor
A leader who answers for the pod by name. The sponsor owns the outcome, approves what the agents are allowed to do, and is the escalation path when one gets something wrong. A pod without a sponsor is a collection of unowned automations, and unowned is the failure mode that kills most agent programs.
The builder
The person who can ship agents: an engineer or a technical operator who turns the expert's knowledge into working automation. Builders are the scarce resource. A pod's build capacity is set by its builders, which is why a pod with a backlog and no builder is just a queue.
The domain expert
The person who actually does the work today. Uber's CTO was specific about why this role is not optional: “You can't automate them effectively by looking at process diagrams or documentation. You have to understand how the work actually gets done.” The expert is how the pod finds the real workflow instead of the documented one.
The agents
The agents are members, not tools. Each has a name, an owner, and a place on the pod's roster; its cost and its output are part of the pod's scoreboard. This is where pods differ from consulting models that staff three humans wielding AI: in a pod, the agents themselves are headcount you manage.
How a pod works: the first ten days
Uber's pods ran a ten-day shape, and it is executable without buying anything. Days one and two: shadow the expert, document every step, build intuition. Day three: prioritize opportunities by scale, repetition, business impact, and data availability. Days four and five: build a working agent alongside the person who does the job. Days six through nine: validate with several others doing the same work. Day ten: ship.
Once a pod is standing, HubSpot's Flanagan runs an ongoing loop: observe the workflows, identify the highest-impact opportunities, build with the expert, deploy to a small group, expand, enable, and measure. His warning about the step everyone skips is the most useful sentence available on adoption: “Building it is maybe 40% of the job. Getting adoption is the other 60%.”
The two meanings of “agentic pod”
The term currently means two incompatible things, and most pages using it do not know the other sense exists.
Sense one: people plus agents. The sense defined on this page. Uber, HubSpot, BCG, and every services firm selling pod capacity mean a unit that contains humans: an expert, a builder, a sponsor, working with agents.
Sense two: agents only. A cluster of software agents coordinating with each other, no human inside the unit. Daniel Vaughan's five-role pipeline is the clearest example: Orchestrator, Planner, Implementer, Reviewer, Validator, all software. “The Orchestrator is the only role that talks directly to the user. Every other role is internal to the pipeline.” QualiZeal describes similar agents-only quality pods for software testing. This is a real pattern, honestly used.
But it is a multi-agent architecture, and architecture already has better names for it. The person who popularized “agentic pod” put a human domain expert inside the unit by definition, and that pairing is the reason his pods worked. When institutions describe agents-only clusters, they reach for other words. A pod, in the sense that is spreading through operating companies, has people in it.
What an agentic pod is not
Not a department
A department groups people by function and owns headcount. A pod crosses functions and owns one workflow, end to end. Departments answer “who do you report to”; pods answer “who automates this work”.
Not a squad or a project team
Project teams disband when the deliverable ships, and the knowledge disperses with them. A pod is permanent. It keeps the workflow it automated, and every agent it ships makes the next one cheaper.
Not a pod you buy
Globant sells AI Pods by monthly subscription: “Clients can subscribe monthly to an AI Pod, and each subscription includes a token-based metered capacity,” per their launch release. Wizeline, Infosys, and HatchWorks sell adjacent delivery units. Those are real services with real results, and they are a different thing: a pod staffed by a vendor's people, priced by the contract. When it ends, the knowledge leaves with them.
| Agentic pod | Department | Project team | A pod you buy | |
|---|---|---|---|---|
| Who is in it | People and AI agents: sponsor, builder, expert | People, grouped by function | People, assembled for a deliverable | A vendor's staff and their agents |
| How long it lasts | Permanent | Permanent | Until the project ships | Until the contract ends |
| What it owns | One workflow, end to end | A function and its headcount | A deliverable | A statement of work |
| Who is accountable | The sponsor, by name | The department head | The project lead | The vendor's account lead |
| How it is measured | Spend, savings, capacity, backlog | Budget and headcount | Scope and schedule | Hours or outcomes billed |
| What happens to what it learns | Stays, and compounds into more agents | Stays, in people | Disperses when the team disbands | Leaves with the vendor |
A Kubernetes pod is a different thing entirely: a group of containers running together on a cluster. Kubernetes had to disambiguate its pod from the whales. We are disambiguating ours from Kubernetes. The two words are unrelated.
What a pod is measured on
Every firm selling pods reports outcome numbers from its own engagements. What nobody publishes is what you should measure on a pod you run. Four numbers, whatever tooling you use:
Spend. What the pod's agents cost this month, metered, not estimated. Savings. What the workflow cost before the pod automated it; the before and after is the pod's reason to exist. Capacity. How much the pod can take on, expressed as a date, because “2 of 4 build slots used, next opens Aug 11” is answerable and “are we at capacity” is not. Backlog. What is queued and waiting, which combined with capacity tells you whether the constraint is ideas or builders.
A pod that cannot produce those four numbers is a project with a nice name. A pod that can is a managed unit of the business, and it can be compared, funded, and fixed like one.
Where pods fail
A pod with a backlog and no builder is a queue. The most common failure. Work piles up behind the one scarce role, and the pod reads busy while shipping nothing. The fix is supply: add a builder or hire one.
A pod with no named sponsor is ownerless. As Gregory Scott Henson puts it: “An agent with no owner is not an agent. It is a science project.”
A pod run as a project disbands, and the workflow knowledge leaves with it. Permanence is not a nice-to-have; it is what makes the second agent cheaper than the first.
A pod measured on activity produces demos. Runs, tokens, and messages are motion. The scoreboard is spend against savings on a workflow someone owns.
Do agentic pods actually work?
The strongest numbers and the strongest objection both come from Uber, and both are public. The numbers, from the CTO's July post: capital allocation across 150 cities went from 15 hours to 30 minutes. Financial pacing reports went from two days to ten minutes. Marketing web quality assurance went from two weeks to fifty minutes. A support process that had accumulated 9,000 manually built workflows was replaced with self-service automation.
The objection, two months earlier: Uber's President and COO Andrew Macdonald, on the Rapid Response podcast in May 2026, said it is “very hard to draw a line” from the company's AI productivity metrics to “ actually producing 25 percent more useful consumer features,” and that “that trade becomes harder to justify because it's not free. AI is not free.”
Both statements are true, and the gap between them is a measurement problem. At company level, the line from AI activity to shipped value runs through too many teams to draw. At pod level it is short: a pod owns one workflow, and a workflow has a before and an after. Two days to ten minutes is not a company-wide productivity claim. It is one pod's scoreboard. That is the strongest argument for the structure: it is the smallest unit you can actually put a cost, a saving, and a date on.
Who runs agentic pods today
Running them internally: Uber, sixteen pods across sixteen functions. HubSpot, where Flanagan runs the model across go-to-market teams. BCG describes agentic-marketer pods of three to five people, each working with agents, in its agent-native marketing model. And McKinsey's technology practice reports early implementations where teams of eight to twelve give way to smaller pods of highly skilled professionals supervising agent-driven execution.
Selling them: Wizeline, Globant, Infosys, and HatchWorks all sell pod-shaped delivery capacity. Different transaction, same shape: the market converged on the unit from both sides.
And GitHub and Microsoft ship the same idea under another name: Squad, “human-led agentic teams,” with the best sentence anyone has written about the shape: the system “can delegate, parallelize, and preserve context. It shouldn't launder accountability.”
Running pods as software
Everything above works with a spreadsheet and discipline. What a platform adds is the scoreboard nobody wants to maintain by hand: live spend and savings per pod, capacity computed from the actual builders and experts on it, and agents that keep their owner and reporting line as the org changes. This is what a pod looks like in SuperOrgs:
Support Pod
2 slots free
Fin, Triage, Scout
The sponsor has the ring. The agents have faces and owners. The capacity line is a date. When a pod stalls with no builder, the card says so and offers the two fixes: add one from your org, or hire one through the platform.
Frequently asked questions
What is the difference between an agentic pod and a squad?
A squad is a team shape from agile software development, usually assembled around a deliverable. An agentic pod is permanent, includes AI agents as members, and owns a business workflow end to end. GitHub and Microsoft ship a similar humans-plus-agents idea under the name Squad; the structure is close, the name is different.
How many people are in an agentic pod?
Small. Uber ran one AI-proficient engineer paired with one domain expert per pod. BCG describes marketing pods of three to five people, each working with agents. Past six or seven people you have a team with a backlog, not a pod with an owner.
Are AI agents members of the pod, or tools the people use?
Members. Each agent has a name, an owner, and a place in the pod's roster, and its cost and output are part of the pod's scoreboard. That is the difference between a pod and a team that happens to use AI tools.
How long does a pod take to ship something?
Uber's pods shipped their first working agent in ten days: two days shadowing the expert, one day prioritizing, two days building alongside the person who does the work, four days validating with others, ship on day ten. Two weeks is a reasonable first cycle.
Who is accountable when an agent in a pod gets something wrong?
The sponsor, by name. Every agent in a pod has a human owner and a reporting line, so a mistake has an escalation path, not a shrug. A pod without a named sponsor is a collection of unowned automations.
Is an agentic pod the same as a Kubernetes pod?
No. A Kubernetes pod is a group of containers running together on a cluster, the smallest deployable unit in container infrastructure. An agentic pod is an organizational unit of people and AI agents. The two words are unrelated.
Can I buy an agentic pod?
You can buy pod-shaped delivery capacity from firms like Globant, Wizeline, Infosys, and HatchWorks, staffed by their people and their agents. What you cannot buy is the compounding: when the contract ends, the workflow knowledge leaves with the vendor. A pod you run keeps it.
Do agentic pods actually work?
Uber's numbers say yes at the workflow level: financial reports went from two days to ten minutes. Uber's own COO says the company-level line to shipped customer value is still hard to draw. Both are true. The honest answer is that pods work where you measure them, which is the workflow.
How do you measure an agentic pod?
Four numbers: spend (what the pod's agents cost this month), savings (what the workflow cost before), capacity (how much the pod can build, expressed as a date), and backlog (what is queued). If a pod cannot produce those four, it is a project with a nice name.
Do I need software to run agentic pods?
No. Uber ran sixteen pods with embedded engineers and a spreadsheet discipline. What gets hard at scale is the scoreboard: live spend and savings per pod, capacity dates, and agents that keep their owner as the org changes. That bookkeeping is what platforms are for.