Designing a Workspace the Agent Can Run

Six properties of a delegable workspace

Work is where the agent can see it. Tasks in Kavanah or in a connected tool, not in a spreadsheet nobody linked.

Ownership is unambiguous. One assignee per task, every task. Work owned by 'the team' is work the agent cannot route, chase or report on.

Outcomes are checkable. Task titles that state what done means. An agent cannot verify 'improve onboarding' and neither can you.

The boundary is written. A charter with a real Negation, so refusing is something the system does rather than something you have to catch.

The calendar is connected. Almost every operational decision touches time, and an agent without a calendar is guessing at the most binding constraint you have.

And context is durable. Decisions on the task, documentation in the knowledge base, standing facts on profiles — not in a chat thread that scrolled away.

A workspace with all six can delegate a lot. One missing any of them will hit a ceiling that no configuration change lifts.

What makes a workspace undelegable

Three things, in order of how often they are the actual cause.

The catch-all project. When everything is in one project, the agent cannot use project context to route, and neither the charter nor access control can be meaningfully scoped. Every question becomes a question about everything.

Due dates that do not mean commitment. A board where half the dates are fictional makes every time-based capability — chasing, prioritizing, planning, overdue automation — generate noise. And noise is worse than absence, because people act on it once and then stop.

Conversation that happens elsewhere. If the commitments are made in a channel Kavanah cannot see and never get written down, the agent is reasoning from a partial record and will confidently tell you about a world that is missing half its decisions.

All three are fixable and none is fixed by a setting.

Sequencing the delegation

Start with work that is reversible, internal, and verifiable.

First: reading and summarizing. Zero risk, immediate value, and it teaches you what the agent can see.

Second: drafting. Emails, task breakdowns, reports. You review everything, so a mistake costs you a read.

Third: internal writes. Creating and updating tasks, moving work, commenting. Reversible via undo, visible in activity trails.

Fourth: recurring internal work. Scheduled runs producing something a person acts on.

Fifth, and only after the ledger says the previous four are reliable: external actions, with the approval step kept in place permanently for the irreversible categories.

The teams that get this wrong usually jump from the first to the fifth, get a bad outcome, and conclude the agent cannot be trusted — when what actually happened is that they skipped the evidence.

The compounding effect

Each of the six properties makes the others more valuable, which is why the returns are non-linear.

Unambiguous ownership plus a connected calendar makes capacity real, which makes planning real, which makes the estimates worth calibrating, which makes the forecasts worth acting on.

A written charter plus durable context makes refusals explicable, which makes people trust the refusals, which makes the boundary hold.

So the practical advice is not to optimize one thing hard. It is to get all six to a decent standard, because a workspace that is good at five and absent on the sixth performs like a workspace that is absent on all of them for anything that touches the missing one.