Waldo

The session is a work log. The outcome is the product. When I was building and operating more than 30 production agent instances at Atlan, the difficult part was rarely getting an agent to produce something. It was remembering why each session existed, moving context between tools, catching a waiting decision, and checking whether the result actually solved the original problem.

Suyash felt the same pressure from another direction while running a design studio and training for an Ironman. His work, commitments, routines, and health lived in tools that never understood how those things affected one another. The software could show more information. It could not decide what mattered now, what could wait, or what no longer deserved to be carried.

Delegation has its own cognitive cost: deciding what to hand off, briefing it, supervising it, recovering from failure, and judging whether the risk was worth it. When that cost exceeds the work avoided, the agent has not really reduced the person’s burden.

The person delegated a problem, not a transcript—and should not inherit a second job stitching every result back into life. AI can do more work than ever. It should not leave you with more to carry.

Give Waldo what you need to become true—not a checklist of every step it should take. Waldo should keep hold of that desired result, coordinate the agents, tools, services, and people contributing to it, and return when judgment, permission, or a change of plan belongs to you.

A responsibility survives the latest task, agent run, interface, and failed attempt. Routine progress should not become another feed to supervise; what remains should not disappear merely because one agent stopped.

Running many agents taught me that starting them is not the hardest part. The hard part is deciding what they may see, what they may change, whether they acted once or twice, what actually became true, and what still belongs to the person.

Waldo is being designed to absorb that delegation tax: compose the permitted context, brief specialist agents, supervise bounded work, recover when a path fails, verify what became true, and return only the judgment or permission that belongs to the person.

The missing layer is agent governance. Waldo is being designed as the owner-side layer between a person and every model, specialist agent, tool, connector, service, or future machine acting on their behalf. It should make delegation useful without letting the machinery grant itself authority or decide that the person’s responsibility is closed.

Kennel and the other working foundations let us test parts of this today. End-to-end cross-surface governance is the target architecture, not a capability we claim as shipped. Kennel and other surfaces can propose work; Waldo’s online governed backend will decide what may become canonical or consequential.

The person must remain able to accept, repair, reopen, defer, transfer, change, or release the responsibility. More machine action should expand personal agency, not replace it.

Today we have three working foundations that we use internally: Kennel on macOS, Waldo on mobile, and the durable backend and agent harness underneath them. Kennel has the strongest bounded acceptance evidence: attributable Codex sessions, conversation history, live processing state, same-task continuation, first-message handling, and archive cleanup.

Kennel is the first home, not the whole vision. It is Waldo’s initial wedge for people already coordinating several agents, giving that work one calm place to land—showing what the agent reports, what evidence supports, what needs judgment, and what remains unresolved without making the transcript the primary unit of value.

The target product is one Waldo carrying responsibilities across work and life. Production integrations, broad provider coverage, automatic artifact verification, and a complete cross-surface responsibility and governance experience are still being built. Today, the evidence is internal and foundation-level rather than proof of the complete integrated experience.

Your life and your agents should not belong to two different systems. Personal assistants and work orchestrators have evolved as separate products, even though the user is the same person. Calendars, messages, reminders, commitments, and daily administration should not belong to a different identity from the sessions, runtimes, tools, budgets, and policies involved in getting work done.

I don’t believe one model or interface will own our entire digital life. People will use many models, specialist agents, tools, services, and devices. Waldo is being designed to join personal assistance and work orchestration through the same Outcome, authority, evidence, and continuity contracts.

An organization may own some infrastructure. The individual should own the continuing relationship. Models, tools, employers, and surfaces can change; the context a person chooses to share, their permissions, corrections, responsibility history, and unresolved work should remain with them.

Waldo may appear as mobile, a judgment in Kennel, messaging, voice, or eventually a physical form. Those are presences of one user-owned personal agent—not disconnected assistants that make the person rebuild context every time. One agent. Many presences. Still yours.

An agent waiting five minutes is inexpensive. A person reconstructing context across five agents, reviewing unverified changes, and finding the correct terminal is expensive. Infinite machine capacity does not create infinite human attention.

Burnout is not a feature category. It is a product constraint: Waldo should not make people supervise more software, monitor more feeds, or remain permanently available. It should carry routine responsibility quietly and return only when timing, consequence, or authority belongs to the person.

Success means less mental reassembly, fewer silently decaying commitments, a realistic next action when capacity changes, and permission to decide that enough is enough.

ChatGPT or Claude can answer what you ask. I want Waldo to understand why you need it, when it matters, what it affects, and whether it was actually resolved. The current prompt is only one fragment of a person’s priorities, relationships, boundaries, capacity, corrections, and commitments.

Life is already distributed across calendars, messages, files, health systems, models, tools, and other people. Waldo should help carry it forward without making the person rebuild themselves—or become the integration layer—every time the interface changes.

The chatbox made intelligence available. It cannot be the whole interface for asynchronous work, and ordinary people should not have to become natural-language programmers or agent managers to benefit. The next layer is continuity, timing, permission, evidence, and closure: proactive enough to prepare what matters, but never presumptive about consequential action.

More execution should never mean less agency. We are building Waldo’s working foundations and target architecture around a human rule: continuity must be inspectable, correction easy, and authority fail closed.

Waldo should learn from what a person says, the corrections they make, and outcomes they verify. Explicit self-knowledge should outrank behavioral inference; activity should never become a hidden personality score.

I see three connected pressures, but I do not treat them as three equal markets. AI-output overload is the customer problem we can attack now through Kennel. Burnout and finite human capacity are the constitution for how Waldo should behave. Physical AI is the expansion horizon where the same questions of permission, evidence, interruption, and recovery become more consequential.

The confidence is different too: the overload and burnout problems are already visible; the physical-world tailwind is strong, but Waldo has not yet validated a hardware product or customer wedge there. The wider curve is consumer: capable intelligence is becoming cheap enough to move agents from specialist tools into everyday products.

Steve Jobs and Steve Wozniak helped turn computers from something hobbyists operated into something ordinary people could make part of their lives. I see agents at the same interface transition: the capability exists, but using it still asks people to think like operators.

A line Jobs wrote about the Macintosh stays with me: “It’s our job to make complex technology easy to use and fun to use.” Waldo is my attempt to do that for agents without hiding intent, consequence, control, or who remains responsible.

I keep returning to bodies for AI: physical forms people would actually welcome into daily life—a desk object, wearable, home device, vehicle, or small robot. The same Waldo should inhabit each of them, carrying one identity and permission system instead of making every object another disconnected assistant.

The physical-AI tailwind is strong, but this is not a current Waldo hardware program and we have low current customer validation for it. Software comes first because identity, correction, permission, evidence, interruption, revocation, and recovery must work before a personal agent is trusted with sensors, movement, or physical authority.

Health and body data matter to me as foundational, permissioned life context; they are not Waldo’s product category. The form may change. The person it works for should not.