Overview
The conversation argues that the next phase of AI will be defined less by larger models than by the systems that let existing capabilities reach users. Kevin Scott describes a “capability overhang”: model reasoning has advanced faster than products have learned to exploit it. Useful agents therefore need durable memory, access to diverse information and tools, permission to act, and open protocols that connect them to external systems. He presents MCP and NLWeb as potential foundational layers of an agentic web, analogous in function—not maturity—to HTTP and HTML. This openness creates security challenges, but Scott believes identity, entitlement discovery, explicit user authorization, and context-aware security agents can support both permissionless innovation and robust protection. On software development, he rejects the idea that agents inherently destroy craft. Makers will continue choosing tools according to whether they value process, outcome, or some mixture of both. He expects many specialized agents rather than one universal system, with differentiation driven primarily by superior understanding of specific user problems. Within a year, he predicts cheaper and more capable technology, increasingly ambitious agent deployments, and a transition toward asynchronous delegation in which agents coordinate multiple systems before returning completed work or requesting human action.
Sections
Higher-Order Insights
Implications that emerge from the combined discussion of capability, infrastructure, security, and product design.
- The agent race is becoming an integration race. As foundation-model capability becomes broadly available, durable memory, tool access, permission systems, and workflow design determine how much intelligence users can actually employ.
- Open protocols could weaken platform gatekeepers by letting creators connect directly with users through agents, but only if participation also makes economic sense for capability and information providers.
- Security may evolve from static application boundaries toward delegated, context-sensitive risk management. An agent with access to multiple communication channels could potentially detect suspicious behavior by triangulating signals that isolated services cannot see.
- Craft is not defined by refusing automation. It is expressed through intentional choices about which parts of creation deserve direct attention and which can be delegated in service of the desired outcome.
Technical Architecture and Security Details
Concrete infrastructure concepts proposed for useful, interoperable, and governable agents.
- Agentic memory must remain coherent beyond a single transaction so agents can continue increasingly complicated delegated tasks without restarting from zero.
- MCP is presented as a simple, open interoperability protocol through which agents can connect to tools, APIs, information sources, and enterprise systems.
- The proposed authorization flow gives an agent an identity, binds it to the person it represents, discovers the systems and entitlements required for a task, and asks the user for the necessary permissions.
- Enterprise administrators need a separate policy layer that determines whether requested agent actions are permitted within each managed system.
- Asynchronous agents will need to make multiple external calls, wait for long-running operations, integrate responses, iterate, and eventually return control to the user.
Predictions for the Next Phase of Agents
Forecasts Scott makes about capability, adoption, product diversity, and interaction patterns.
- Developers and companies that postpone experimentation because agents are currently slightly too expensive or insufficiently capable will fall significantly behind.
- Agent systems will become cheaper and more capable each year, weakening arguments that builders should wait for the technology to mature.
- People will apply agents to increasingly ambitious problems as planning improves and the agentic web becomes more complete.
- Software engineering will support many competing agents with different product choices and specializations rather than converging on one tool used by every developer.
- Agent interaction will shift from immediate prompt-response exchanges toward asynchronous assignments that may run for a substantial period before returning results or requesting human action.
Memorable Quotes
Statements that capture the interview's central arguments in the speakers' original words.
- I've been talking a lot lately about this thing called the capability overhang.
- if agents are going to be useful, they have to take action on your behalf.
- You ship your org chart.
- If you value the process more than you value the outcome, like sometimes you'll make different decisions than people who value the outcome more than the process.
- be curious, try stuff. Like and if it works for you, use it, and if it doesn't, don't.