Last Thursday at Gemini at Work, Google Cloud CEO Thomas Kurian unveiled a 'universal agent' for the workplace: it writes and runs code, generates media, plugs into Gmail, Docs, Drive and Sheets as well as Slack and Microsoft 365, and — this is the part that should make every operator sit up — creates its own 'temporary, job-specific agents, each with their own identity,' that can 'work together to accomplish goals over several days.' It builds memory. You never have to re-brief it. It is, in Google's words, a co-worker that carries out tasks autonomously. Day one of that is magical. Day three is where every autonomy story is actually decided, and the launch coverage did not spend a sentence there.
1. Day One: It Does the Thing
The multi-day, multi-agent pitch is real and it is the right direction. The work that matters in a business is rarely a single prompt; it is 'get the Q4 forecast reconciled with the pipeline, chase the three accounts that went quiet, and have a draft for the board by Friday' — a goal that spans days, tools and people. An agent that can hold that goal, decompose it, and keep going without a human nudging it every forty minutes is the difference between an assistant and an employee. We have been building exactly that shape of worker for a year: scheduled autonomous runs with a standing instruction, a memory of what was done, and a report back when it is finished.
2. Day Two: It Does Something You Didn't Expect
By day two a long-running agent has made dozens of decisions you never saw. It chose whom to email. It decided a stale thread was worth reviving. It spun up a sub-agent with its own identity to go do something. Gartner's warning about agentic AI is blunt on this point: current models 'don't have the maturity and agency to autonomously achieve complex business goals or follow nuanced instructions over time.' Multiply a small misreading by three days of unattended work and you get a very confident, very wrong result delivered on time. Which raises the only question that matters: between day one and day three, what stopped it?
3. Day Three: Who Approved That, and What Did It Cost?
Here is what a multi-day agent needs that a demo never shows. A policy layer that parks external sends, payments and deletions for a human to approve rather than performing them on its own authority. A per-day token ceiling so a loop that goes sideways on Tuesday night cannot spend the quarter's AI budget by Wednesday morning. A ledger of every action — what tool, what target, what outcome — so 'what did it do while I was asleep?' has an answer that is not a vibe. And an undo for the actions that can be reversed. Deloitte found only 21% of companies have a mature governance model for autonomous agents. The other 79% are about to find out what day three looks like.
4. How Kavanah Answers Day Three
Every AI Employee in Kavanah runs under a scope you assigned, a safe mode that strips risky tools from unattended runs, and an action-policy engine that can allow, cap, park or block a category of action — so the agent that drafts the board update cannot also send it to the board without a human clicking approve. Headless runs have a daily token budget that stops the model before it runs rather than after the bill arrives, and every model call is billed to a named payer at the price of the model that ran. Every action lands in an audit ledger with an undo where one exists. An agent can absolutely work on a goal for days here. It just cannot do anything on day three that you would not have let it do on day one.
5. The Test to Run Before You Turn Autonomy On
Whatever multi-day agent you are about to deploy — Google's, ours, anyone's — give it a goal that will take seventy-two hours and then answer three questions before you press start. What is the complete list of things it is allowed to do without asking? What is the most it can spend before it is stopped? And on Friday morning, where do I read the full account of what it did? If any of the three has no answer, you are not deploying an employee. You are releasing a very capable stranger into your company and hoping.
Agents that work for days are the future of work. Agents that work for days with no scope, no budget and no ledger are the future of incident reports.
