Open Devin
Lesson 6 · Video · 4 min

Agents as Workforce

Optional reading: go deeper

So far the focus has been a single agent on a single task. Now zoom out to the org level: what changes when agents are part of your capacity, not just a tool you reach for.

The shift is in how you allocate work. A tool waits to be picked up; you only use it while you are already working on something. Workforce is different: you assign it work from the backlog, and it runs while you do something else.

Mental model Implication
"AI is a tool" I use it when I'm already working on something. It accelerates my hands.
"AI is a team member" I assign it work from the backlog. It runs on its own while I do something else.

Once you make that shift, new questions surface that you would never ask of a tool:

  • What is stuck in the backlog because no one has capacity?
  • What is well-defined enough that an agent could execute it from a clear prompt?
  • How much of every sprint goes to routine maintenance that could run on a schedule instead?
  • If we parallelized a campaign across twenty agents, what backlog could we clear?

Three tiers of work

In a team that includes agents, work distributes across three tiers. These are not a hierarchy; they are three modes of interaction.

Tier Owns Examples
Human engineers Judgment Architecture and system design, novel problem-solving, product decisions, cross-team coordination, exception handling, final code-review authority
Cloud agents Execution Well-defined tasks, repetitive work at scale (migrations, upgrades), event-driven responses (CI fixes, security triage), scheduled maintenance, investigation, test generation
Local assistants Acceleration Code completion, real-time pair programming, quick refactors, inline explanations while you are in active focus

A single piece of work can touch all three: use a local assistant to explore a design approach, delegate the implementation to a cloud agent, then review the PR and make the architectural call yourself.

How agents get assigned work

Delegation is not always one prompt to one agent. A few common patterns:

  • Single agent: one task, one agent, one PR. The default for most work.
  • Parent-child: a campaign prompt goes to a parent agent that scopes the work and spawns N child agents, each on its own VM and branch. A failure in one child does not affect the others.
  • Event-driven: an agent reacts to a signal from your toolchain (a CI failure, a security finding).
  • Scheduled: a recurring session does routine hygiene without anyone kicking it off.
Back to course← Previous lesson
© 2026 Cognition AI, Inc.Made with by the Cognition team