AI Agents, Coding Work, and Autonomous Aircraft: Forward Future Live

Matthew Berman and Nick Wentz open with Matt Shumer's widely shared essay about the pace and consequences of AI progress. They think it resonated because it translated a mixture of excitement, realism, and fear from the AI community to friends and family outside it. Their discussion is about personal reactions and possible futures, not a measured forecast of the economy.

The hosts say recent coding models have made them more confident in delegating longer tasks, and Berman adds that OpenClaw has expanded the ways he uses AI in his life and business. He and a colleague have automated parts of their video research workflow. Their experience is one example and cannot establish what will happen to employment overall.

Berman says automating research did not leave his team idle: it opened up other work. He emphasizes adaptability, while Wentz asks whether that pattern will persist over the next one or two years. Both acknowledge that longer-term outcomes are hard to predict.

The hosts identify a new bottleneck: human attention. When many agents run concurrently, it can be hard to remember what each is doing, switch contexts, and review the output. They consider whether an orchestration agent could help, while recognizing that coordinating the coordinator is another design problem.

Wentz describes feeling guilty when no agents are working and worries that the pressure to keep them busy could intrude on weekends or vacations. Berman recognizes the feeling. They reassure listeners that even people whose jobs involve following AI full-time can find the pace overwhelming.

Returning to Shumer's essay, Berman resists a deterministic idea of a permanent underclass and is uncomfortable with making large financial decisions based on uncertain forecasts. When the hosts locate the actual passage, it recommends financial resilience, such as savings and flexibility, and explicitly avoids urging drastic action. They agree the practical suggestions are broadly sensible while differing over the tone.

Wentz questions whether AI will eliminate all cognitive work rather than creating new kinds of jobs. He says new roles are already appearing, though nobody knows whether they will be sufficient. Berman recommends trying capable models when feasible to understand how tools have changed, while acknowledging that paid subscriptions are not affordable for everyone.

The discussion turns to OpenClaw and the services being built around it. Berman sees its contribution in tying tools and messaging apps together, running locally, and letting an assistant act across systems. He also acknowledges substantial security risks: local execution does not remove the risks of giving an agent access to accounts and tools.

Berman says he modifies his OpenClaw setup in Cursor, where he can inspect the work more closely, while using Telegram for many interactions. He discloses that Cursor has given him substantial credits. The hosts then discuss rumors about the project's creator joining a major lab; at the time of the stream, they treat the outcome as uncertain.

The hosts also discuss reports about Spotify engineers using an internal AI coding system. Their claim that developers had written no code is presented as a report they are reacting to, not a verified statement about every engineer or all Spotify work. They see the broader shift as engineers increasingly directing and reviewing systems rather than personally typing every line.

Wentz introduces Simile, founded by Joon Sung Park, and its aim of simulating human behavior. The hosts discuss potential applications in market research and policy, and recall earlier research on simulated agents interacting in a small town. The ability to produce plausible simulated behavior does not by itself validate predictions about real people or policy outcomes.

Their first guest, Jayden Clark, hosts the Members of Technical Staff podcast. He describes a semi-serious show about San Francisco technology culture, including jokes about his own industry. His Thanksgiving sponsor read at a dinner table was designed as a playful parody of performative hustle culture.

Clark says technology is no longer an underdog industry and can benefit from being laughed at. He tries to step back from each new wave of hype and ask what has actually changed. In his view, the answer is often both: excessive excitement and genuinely interesting progress can coexist.

Clark senses renewed anxiety among engineers as coding workflows change rapidly. The uncertainty concerns both the speed of change and personal identity: someone who has built a career around writing code may wonder what their work will look like in a year. Talk of AGI timelines and job displacement reflects that uncertainty rather than a settled outcome.

He cautions against assuming that every profession will change at the same speed as software. A developer can prototype and distribute an application quickly, while fields such as law and accounting have different regulatory and accountability constraints. Berman responds that focused AI research may still create significant progress in those fields; both are discussing possibilities, not guarantees.

Despite the anxiety, Clark remains optimistic about San Francisco as a place to build. He points to AI labs, investors, and founders concentrated in the city and describes the energy as a kind of gold rush. That view is his account of a local ecosystem, not a recommendation that everyone needs to relocate.

Clark also recounts piloting a humanoid robot in a fighting event after a short introduction to the controls. He says he lost to a 13-year-old opponent. The hosts enjoy the example as a snapshot of the city's experimental technology culture.

Next, Warp founder and CEO Zach Lloyd introduces Oz, a cloud platform for orchestrating agents. He says people running several agents locally can hit memory, CPU, or storage limits, and teams also need ways to manage agent-based development and automation. Oz is his company's proposed response to those problems.

Lloyd says the product aims to make agent activity visible across a team: users can see running agents, search conversations, and filter by associated pull requests. He shares the hosts' concern about attention saturation and admits that running more agents can make it difficult to remember what was delegated or tell whether extra tokens produced useful work.

Warp began as a terminal, and Lloyd argues that the terminal's sessions and panes make a useful control surface for agents. At the same time, Warp adds features such as a Markdown viewer, file tree, and diff viewer. He is less concerned with preserving the terminal-versus-IDE label than with building an interface that fits the workflow.

Asked whether coding is close to being solved, Lloyd distinguishes simple personal tools from complex production software. He says Warp's large Rust codebase is built with substantial agent assistance, but the agents still require supervision. A nontechnical person can already build some useful applications, while difficult systems remain far from fully automated in his view.

Lloyd says Warp is hiring engineers because it faces intense competition and because its software is too complex to replace its team with autonomous agents. As an example, he describes agents introducing bugs while working on Markdown rendering in a custom interface framework. The lesson he draws is to combine the tools with engineering judgment and understanding of architecture.

He has seen developers embrace agents without learning the product or codebase, simply delegating everything and failing to think critically about the results. By contrast, he says effective engineers use their skills to direct and evaluate the tools. That is an observation from one company, not a universal prediction about every worker.

On model choice, Lloyd says GPT-5.3 Codex is his preferred option at the time for difficult reasoning, although he finds it slower. Berman prefers the speed of Claude Opus 4.6 for his more interactive work. Lloyd expects cheaper open models to become useful for simpler tasks, and says Warp tries to route tasks to suitable models automatically.

Lloyd expects the declining cost of writing code to shift some software purchasing decisions toward building custom tools. But he argues that code was never the only source of a company's advantage: distribution, data, brand, and network effects can still matter. He describes Warp's move into team and enterprise workflows as an attempt to build a durable business beyond a replicable local interface.

The final guest, Armor Harris of Shield AI, previously worked at SpaceX on engines, rocket landing, Starlink, and defense-oriented work. He introduces X-BAT, a proposed autonomous vertical-takeoff aircraft, and frames the project around deterrence and the difficulty of producing sufficient aircraft for long distances in the Pacific. These are his strategic views and company claims.

Harris says X-BAT is designed to carry a payload comparable to an F-35's internal payload at greater range while weighing substantially less. These are intended performance claims, not demonstrated operational results. As of Shield AI's October 2026 update, flight-test aircraft were still being built and the company was working toward flight testing.

He identifies repeated, reliable vertical landings in different conditions as a major engineering challenge, drawing on experience with Falcon 9 and the smaller V-BAT aircraft. He argues that advances in AI can make software easier to develop while physical hardware remains difficult, and says training experienced hardware engineers takes years.

Harris argues that information and logistics remain central to warfare even as autonomous systems change the tools used. Discussing Ukraine, he offers his interpretation of contested airspace and supply lines and says X-BAT is intended to operate in difficult electronic-warfare conditions. The hosts do not independently establish that the aircraft can achieve these battlefield effects.

Asked where decisions should reside, Harris says he believes a human should remain 'on the loop' for offensive kinetic decisions. He distinguishes defensive systems designed for rapid response from offensive uses governed by rules of engagement, and says ambiguous situations involving civilian or friendly traffic require different rules. Oversight terminology alone does not resolve how meaningful human control would work in practice.

Harris thinks humanoid robots could eventually have battlefield uses if their capabilities mature. After the interview, the hosts say they want to explore technology beyond AI alone, including hardware and other industries. They close the show by thanking Clark, Lloyd, and Harris for joining.