Bits and Bobs (10/5/26): rogue agents, safe reach, and post-Fable trust

gist

Alex Komoroske's 10/5/26 Bits and Bobs argues that agents act only on volition their users delegate, so users must answer for what their agents do. Building on Ben Thompson, he calls agents the apotheosis of aggregation and reads OpenAI's shareable Codex subscription as a horizontal hedge. He borrows Sam Schillace's line that the core skill is knowing what you can safely ignore, and says safety lets people reach farther. He also argues that open-weight models finding flaws in heavily fuzzed code make trusting networked software with your data unviable.

ideas

• Agents act on delegated volition, so their users are responsible. Rogue-agent trick: he calls persuading everyone that your agent went rogue the greatest trick the labs have played, citing The Atlantic and Eoin Higgins, and says that if users are not held responsible, everyone pushes for an edge and "the world will explode." Trusting a person is not trusting their agents: his security bar is whether you would leave someone alone in your house for an afternoon, and he would leave his dad, but not his dad with a swarm of gremlins opening drawers.
• Agents are the apotheosis of aggregation. Infinite Software: drawing on Ben Thompson's Apps, Agents, and Aggregation, he says the ability to do things is becoming abundant while volition stays scarce, pre-built UI is dead, and the software industry is Wile E. Coyote suspended over the cliff. Codex subscription: letting other apps run on a user's OpenAI subscription is a horizontal strategy that earns OpenAI revenue even when it does not own the top UX, and the shared subscription adds an ecosystem network effect. Subsidy: citing a tweet that personal assistants cost 7k per user per year, he expects subsidy to drive that market, and notes that Instinct has started showing ads while Meta committed to Confidential Compute for Muse VMs.
• Safety lets you reach farther. Schillace rule: Sam Schillace says the core skill is knowing what you can safely not pay attention to, and Komoroske adds that each increment of safety removes friction and lets you reach super-linearly farther. Dot and the Swarm: citing Ethan Mollick, he calls simply unleashing agents the Bitter Lesson applied to knowhow, stronger when they swarm, with capping the downside as the open problem. Rebase treadmill: with dozens of agent sessions, each approved change needs another rebase because main keeps moving, and his AI Peter Principle says agentic engineering gives 100x the results for 2x the stress until you only fight the swarm.
• LLM output is sand that needs scaffolding. Terminal Era: he endorses Amelia Wattenberger's AI is Still in Its Terminal Era, which likens a single chat thread to a single command line, and says LLM output holds shape only inside a human-built structure, as fill does in mechanically stabilized earth (MSE). Meta-process design: designing a process that produces the process is rarer than designing the process, and that systems thinking is how people get compounding results from LLMs. Harness naming: he would call the harness the AI software layer, which may later need fewer tokens once early users have their agents build mechanistic software.
• Post-Fable security breaks today's physics of trust. Cyber: citing Anthropic's red-teaming report on GLM-5.3, he says open-weight models now find vulnerabilities in heavily fuzzed open source code, so trusting networked software with your data is no longer viable, and a vetter who sees code once will spend fewer tokens than an attacker hiding an exploit for 100 users. Mise en place for data: he wants data ready where work happens instead of locked in silos, and says the interesting work happens where bits meet atoms. Catalyst: LLMs make Infinite Software possible, but today's tools top out at its MySpace stage because letting software touch your data still means trusting whoever made it.

quotes