Three new cards from three operators. Each lands in a different lane, but the pressure connecting them is the same: AI is changing the economics and the narratives of how work gets done, and the early claims are not holding up the way the headlines suggested.
Theme 1, AI-cited layoffs mask structural corrections
Kevin Indig's H1 2026 halftime report (AI-cited layoffs are mostly a cover story for pandemic overhiring and capital cycle corrections, not evidence of automation-driven job displacement) examines the wave of layoffs attributed to AI and finds the evidence does not support the attribution. The tell is in the reinstatement data: more than 30% of companies that made AI-cited cuts hired back similar roles within months, according to Robert Half survey data. Genuine automation-driven displacement does not produce this pattern. Roles come back when the work was not actually eliminated, only the headcount budget was.
"Despite clear productivity gains from AI, the recent waves of layoffs show no indication of AI being the underlying cause."
The mechanism Indig names is narrative utility. Saying "AI is replacing these roles" is a forward-looking, productivity-framed justification that avoids the admission that a company over-hired during the 2020 to 2022 expansion and is now correcting. The AI narrative is useful because it sounds like strategy rather than mistake. The reinstatement data breaks it.
- Kevin Indig, AI-cited layoffs are mostly a cover story for pandemic overhiring and capital cycle corrections, not evidence of automation-driven job displacement. AI-cited layoffs show reinstatement rates above 30%, indicating the underlying cause is pandemic overhiring correction, not automation-driven displacement.
Theme 2, The terminal multiplexer is the emerging coordination primitive for mixed human-agent teams
Mitchell Hashimoto launched Superlogical on July 29, framing a new company around a single design claim: the terminal multiplexer built for user sessions is the wrong primitive for the agent era. Current terminal multiplexers fail by duplicating state client-to-server; a stateful session per task lets humans and AI agents share the same work context asynchronously captures the founding thesis.
The current failure mode Hashimoto identifies is precise. Existing multiplexers like tmux and screen were designed for single-user remote access. They duplicate everything the user types and constantly reconcile local and remote state. This architecture breaks when an AI agent and a human developer need to share the same execution context: the session is tied to a person, not to the work.
"the common denominator across today's developers, agents, tools, and infrastructure"
Hashimoto's alternative is a stateful session per task, carrying context, data, and history forward independently of who is attached. Any participant, human or agent, can resume from the same point without manual context restoration. The terminal is not being replaced; it is being redesigned around the work unit rather than the user unit.
The finding connects to Tomasz Tunguz's benchmark evidence in The harness moves coding benchmarks more than the model does; third-party harnesses already outperform first-party ones on the same model: if the layer above the model determines coding performance more than the model itself, then the layer below the harness (the execution context where agents and humans share state) becomes the next surface worth engineering carefully.
- Mitchell Hashimoto, Current terminal multiplexers fail by duplicating state client-to-server; a stateful session per task lets humans and AI agents share the same work context asynchronously. The terminal multiplexer, rebuilt around task-scoped persistent sessions rather than user sessions, is the coordination primitive for mixed human-agent teams.
Theme 3, Vibe-coding flips the SaaS build-vs-buy calculation for solo operators
Pieter Levels published his account of canceling every SaaS subscription he paid for and rebuilding each tool himself with AI assistance. Vibe-coding collapses custom tool build cost enough that recurring SaaS subscription costs now exceed one-time build cost for solo operators names the economic inversion.
SaaS pricing has always assumed build cost is high enough to justify recurring fees. A $49/month tool is a good deal when building an equivalent takes a week. That assumption held when custom development required time or money that solo operators did not have. Vibe-coding collapses the build side: Levels rebuilt his full SaaS stack himself, cutting recurring spend to domains, hosting, storage, and AI APIs.
"I cancelled and then vibecoded 100% of my SaaS subscriptions"
The limits of this claim are real. Levels is an outlier by design: decades of solo ship experience, no employees, and exceptional build proficiency. Most operators who attempt the same substitution will underestimate maintenance and debugging costs. The initial build is cheap; the lifetime upkeep is not. But the threshold claim holds directionally for the narrow profile Levels represents: solo or micro-team operators with AI-assisted coding access and single-function tool needs.
- Pieter Levels, Vibe-coding collapses custom tool build cost enough that recurring SaaS subscription costs now exceed one-time build cost for solo operators. Vibe-coding collapses custom tool build cost enough that recurring SaaS subscription costs now exceed one-time build cost for solo operators with strong AI coding access.