From 100 Agents to Two Heroes
TLDR
If every team has an agent and nobody owns adoption, you did not transform. You multiplied pilots. Box’s public AI-first series (part 1, part 2) shows the path: let many people experiment, then consolidate into a few hero agents with real change management. Aaron Levie on CXOTalk 921 (YouTube) explains why coding agents take off while knowledge-work agents stall. The scarce asset is not another model. It is who picks, who builds for scale, who owns adoption, and what “done” means.
What “hero agent” means (and what it does not)
The phrase is fuzzy in most enterprises. People use it for any demo that got applause.
In Box’s language (part 2):
- A big bet is an idea selected for pilot because of potential impact.
- A hero agent is a piloted agent that has proven high impact and then gets leadership-backed scaled adoption: workflow redesign, training, usage targets, and tracking.
Scope that matters: heroes are few, owned, measured, and wired into how work actually runs. Out of scope: agent count as a KPI, one-off sandboxes with ten users, and “we launched AI” without an adoption OKR.
Experimentation is a phase, not the strategy
Part 1 opens with the usual trap: disconnected micro-initiatives, tools everywhere, transformation nowhere. Box did not ban experimentation. It started with policy guidelines rather than a technical lockdown, sandbox access for people closest to the work, and idea submission for people with the pain but not the time to build.
That phase is useful. Without it, central planning guesses. With it, you see which workflows improve and which ideas are toys.
Olivia Nottebohm (Box COO) in part 2: spinning up your own agent can work for smaller tasks. Transforming the business needs intention. Ten people using an agent is not a win. Hundreds of people using two or three agents that reshape the work is.
Peer reply: keep the hopper. Starve the portfolio.
Heroes are an ownership design
Translate Box’s model into operating terms. The scarce assets are:
- Who picks. Functional leaders propose big bets inside central strategy and guardrails. Executive sponsors make trade-offs and say no (part 1).
- Who builds for scale. Local teams can ship single-function agents. Shared Design and Build capacity handles hard integrations so six teams do not rebuild the same account-research agent six times.
- Who owns adoption. AI Managers drive training, feedback, and recognition. Targets look like “every AE uses meeting-prep for every meeting,” not “we shipped 40 agents.”
- What done means. Part 1 splits value into productivity, automation, and net-new capability. Pilots measure those claims. Rollout needs reliability, knowledge pipelines, and governance. Scaled adoption measures use and impact.
This is roles, workflows, decision rights, and controls. Agents are the surface.
Why sprawl feels productive
Overlap. Part 2 describes merging upwards of ten similar ideation agents into a couple of stronger ones. Consolidation buys coaching, data, and integration budget.
Permissions. Connecting agents into systems like Salesforce or GCP is where value appears, and where security gets expensive. Sprawl multiplies auth, audit, and leakage risk. On CXOTalk 921, Levie notes coding agents often inherit an engineer’s codebase access. In the enterprise, agents either lack the right data or get too much of it.
Knowledge work is not coding. Levie’s contrast: engineering is mostly text, technical users, often verifiable with tests. Knowledge work is less verifiable and permission-messy. Copy “everyone spin an agent” from coding into sales or support without redesigning verification and ownership, and you get theatre.
CEO distance. Levie calls it AI psychosis: leaders are far from the last mile. Demos look like automation. Closer in, wrong data pulls and maintenance remain. Humans stay in the loop. Design for that hybrid.
What to do Monday
- Force each function to name two or three priorities. Give sponsors licence to kill the rest.
- Separate propose, decide, build, and adopt. One group owning all four either freezes innovation or ships unowned pilots.
- Redesign the workflow around the hero. Task automation inside an unchanged process dies on a slide.
- Treat non-use as a control failure.
- Run a monthly pilot review and a quarterly big-bet reset (part 2).
- Put budget with the line. Levie argues IT cannot hold the entire AI budget if AI is to matter at scale.
Sources
- Box, AI-First Transformation: Box’s Principles, Strategy, and Execution Framework.
- Box, From 100 Agents to Strategic Big Bets.
- Aaron Levie with Michael Krigsman, CXOTalk episode 921 (YouTube).
Challenge
If you cut to two hero agents per function tomorrow:
- Who owns the adoption OKR by name?
- What workflow changes on Monday, not on a roadmap slide?
- Which twenty agents die this quarter, and who has air cover to kill them?
If you cannot answer those three, you do not have heroes. You have a catalogue.