The work is deciding
what matters.

We believe that as AI gets more capable, more of our work will shift from figuring out how to do something to deciding what is worth doing. Which problems deserve effort? What should happen first? What does a good result actually look like?

Execution will still matter. But when many agents can work in parallel, the scarce resource becomes the direction we give them. A company can do more and still spend its effort on the wrong things.

The bot in the channel

Putting an agent in Slack is an obvious place to start. It meets people where they already work. You ask for something, and it responds.

Now imagine several agents working for several people. Each posts its plan, its progress, its questions, and its results. Each message may be useful on its own. Together, they become another stream somebody has to read, interpret, and coordinate.

We think this creates agent fatigue: people begin to ignore the very systems that are supposed to help them. The problem is the attention those systems require to stay useful.

The supervision trap

Agent speed makes that problem more interesting. Humans and agents operate at different tempos, yet we keep putting their coordination into the same conversation.

Consider a task with several possible approaches. In the time it takes a person to read a proposed plan and respond, a capable agent team might have explored five approaches in parallel, compared them, and revised its plan.

This is the supervision trap: you delegate the work, then spend your time managing the agents doing it. Every plan to approve, update to interpret, and handoff to coordinate becomes another task on your plate.

Making people approve every decomposition or relay information between agents puts human response time in the middle of machine execution. Agents should be able to handle that coordination themselves.

An organization on the other side

Our bet is that agents need an organization of their own: a place to share context, divide work, resolve dependencies, compare results, and revise their approach.

Alongside it sits the human team, where priorities are debated, intent is clarified, and tradeoffs are decided. These two groups need an interface that lets each work at its own pace.

That separation is about protecting attention. People should be able to inspect what happened without having to follow every step as it happens.

Intent is a team decision

In an organization, intent rarely comes from one perfect prompt. One person sets a deadline. Another changes the scope. Someone else knows a constraint that never made it into the brief.

The system needs to maintain a shared understanding of what the team wants. When instructions conflict, it should bring back a clear decision to the right person. When the path is already authorized, agents should keep working.

The human role becomes defining intent, allocating effort, and judging the result. Human involvement remains essential for consequential choices; it should not be required for every operational detail.

Systems shaped around the business

Every organization has its own understanding of the work: the people involved, the records it keeps, the relationships between them, and the rules that govern what happens next. Together, these form its enterprise ontology.

We believe agents should learn that structure and build systems around it. Shared records, actions, and workflows give an agent team a durable way to operate. The interface gives people a way to direct that system and inspect what it produces.

This is what we mean by a dynamic backend. As the work changes, agents adapt the underlying model and behavior alongside the interface. A new rule should carry through to how the system acts, not simply appear in another message.

Make the work easy to verify

Delegation only saves attention if checking the result takes less effort than doing the work yourself. A confident message saying “done” is not enough. People need to see what was produced, what supports it, and whether it meets the intent they started with.

We believe the interface should take shape around the work. A comparison might need a table with sources. A change might need a before-and-after view. A completed action might need a confirmation. Agents should assemble the view that makes the result easiest to inspect.

As the work changes, that view should make clear what is current, what changed, and what still needs checking. Sources and checks should sit alongside the claims they support, including any uncertainty or missing evidence.

People should be able to set direction, let the agents coordinate, and return to verify the result. The interface should make that verification straightforward, without requiring someone to reconstruct the process from a stream of messages.

What we’re building toward

Bandit is our attempt to build that relationship between human teams and agent teams. The interface matters as much as the execution behind it.

Conversations can start the work. Shared memory carries decisions forward. A living artifact shows the current result, its evidence, and what changed. Questions return when human judgment can materially change the outcome.

That is the future of work we want to build toward: people deciding where effort belongs, with an organization of agents capable of carrying it through.