Four cards from three operators. Two themes from the Airbnb CTO. One claim from Mollick that reframes the headcount question entirely. One pricing alert from Willison that any agent builder should have in front of them.
AI rewrites the org chart
Ahmad Al-Dahle appeared on Latent Space on October 2. Airbnb is shipping AI-authored code at 60% of new output, 80% more features year-over-year. That operational reality is what makes his structural prediction blunt:
"If you're a company that organizes by feature development, you're going to struggle in the age of AI."
Feature-organized companies will struggle as AI ships features autonomously; outcome-organized teams set the direction AI cannot replace captures the claim and its mechanism. Feature teams exist to ship features. When an AI plus a codebase knowledge graph can ship a feature in six weeks rather than nine months, the team organized to produce that feature loses its rationale. The unit of organization that survives is the outcome mandate. Direction and judgment about what to build and for whom cannot be automated. Feature execution increasingly can.
This sits alongside a second Al-Dahle claim that is more immediately actionable for engineering teams. Small post-trained models outperform frontier models on narrow latency-sensitive tasks, making multi-model routing the default production architecture documents Airbnb's production routing logic: frontier models for coding (defect cost is high), small specialist models for search (latency dominates). Al-Dahle:
"take really small, really nimble models that are very fast and cheap."
The mechanism is post-training concentration: a small model trained on a narrow production query distribution will beat a frontier model on that distribution precisely because breadth is wasted capacity. Multi-model routing is not a future architecture. It is how Airbnb is running today.
- Ahmad Al-Dahle, Feature-organized companies will struggle as AI ships features autonomously; outcome-organized teams set the direction AI cannot replace. Feature-organized companies face structural obsolescence as AI ships features autonomously; outcome-organized teams set direction AI cannot replace.
- Ahmad Al-Dahle, Small post-trained models outperform frontier models on narrow latency-sensitive tasks, making multi-model routing the default production architecture. Narrow post-trained models outperform frontier models on latency-sensitive tasks; route by error cost, not by benchmark rank.
Cheap coordination changes what gets started
Ethan Mollick's "The Dot and the Swarm" (October 1) contains a second claim beyond the self-organization thesis captured last week. Cheap agent coordination expands the set of tasks organizations attempt rather than simply reducing headcount extracts it.
The efficiency frame for AI agents is: existing work, fewer people, lower cost. Mollick's claim is different. When coordination overhead drops near zero, the barrier that kept speculative and exploratory projects off roadmaps drops with it. Projects that were never worth starting because they were too expensive to staff and coordinate become worth starting.
"When it gets cheap, the list of things worth attempting can grow."
The organizational implication follows: the primary near-term effect of cheap agent coordination is not headcount reduction. It is an expansion of what organizations are willing to try. The right question for a leadership team is not "how many people can we remove?" It is "what projects were we not starting that we should now attempt?"
- Ethan Mollick, Cheap agent coordination expands the set of tasks organizations attempt rather than simply reducing headcount. Cheap agent coordination expands organizational scope; the first-order effect is more attempts, not fewer people.
The Haiku cost cliff
Simon Willison published his analysis of Claude Haiku 5.5 on October 7. The headline number is the 100k-token context boundary. Below it: $0.10 input / $0.50 output per million tokens. Above it: $0.50 input / $2.50 output. That is a 5x jump.
Haiku 5.5 pricing jumps 5x at the 100k-token context boundary, creating a cost cliff that long-running agent loops will cross captures the mechanism agent builders need to internalize. A chat session with a human rarely crosses 100k tokens. An agent loop accumulates context across every tool call, every intermediate step, every iteration. A loop that starts at Haiku 5.5's low base rate can cross the boundary mid-run and pay 5x pricing for the remainder. For comparison, Luna (xAI) holds its low base pricing through 272k tokens before its cliff.
The action is straightforward: model your loop's expected context accumulation profile before committing to Haiku 5.5. If your loops routinely cross 100k tokens, the headline cost is not the operating cost.
- Simon Willison, Haiku 5.5 pricing jumps 5x at the 100k-token context boundary, creating a cost cliff that long-running agent loops will cross. Haiku 5.5's 5x pricing jump at 100k tokens will be crossed by long-running agent loops; model context accumulation before defaulting to the low headline rate.