Tap Notes: Load-Bearing
The 40x subsidy number has been floating around for a while. Today it landed with a date attached — Microsoft canceling Claude Code at the end of this month. The rest of today’s reading is either the vision that makes the cost worth it, or the architecture that makes the math survivable. One piece zooms out to mathematics and asks a question that turns out to be about agents.
A $200/month Anthropic subscription consumes ~$8,000 in actual token costs. Anthropic announced Agent SDK token billing, then paused it one day later when users revolted. The 2024 MIT stat underneath the investment thesis: 77% of tasks are still preferable done by humans at current capability levels.
The current pricing has trained entire engineering organizations to treat AI as near-zero marginal cost. Any correction feels like betrayal rather than math.Post to X
The question worth building into your architecture now: what would you actually pay if the subsidy ended tomorrow?
Anthropic ships a Slack integration where Claude participates in channels. The detail that matters is ambient behavior — Claude proactively flagging relevant info and following up on quiet threads without being tagged. That’s the difference between “tool you summon” and “teammate who reads the room.”
Credibility signal: Anthropic uses this internally for 65% of their code, meaning they’re describing their present, not a future. The scoped identity design — separate instances per channel, sealed memories, no cross-domain leak — is also the architectural pattern that prevents the horror scenario where your AI assistant bleeds context across workspaces. They’ve proven it at scale.
Sakana Fugu — Multi-agent System as a Model
A system that routes across a pool of frontier models, learns which ones perform best per task type, and lets you exclude specific vendors for compliance. The goal: frontier-level performance without single-vendor dependency.
The infrastructure argument is practical — when your system runs on one model API and that vendor has a bad week (pricing, downtime, policy shift), your whole system goes with it. But the finding that stands out comes from a real user: Fugu “holds identity where other models drift.” That’s a failure mode that’s invisible in real time. You only see the accumulated damage after a long session where model substitution has quietly eroded agent consistency. Routing resilience is architecture, not a patch you apply after the fact.
A quote from Jon Udell (via Simon Willison)
The loop is fundamentally human — goals, taste, judgment, veto power. Agents execute within constraints the human sets.Post to X
“Human in the loop” puts the machine at the center and treats human oversight as an interruption. That framing is backwards. Transparency isn’t an afterthought in this model — it’s the mechanism that keeps the human loop intact. Constrained autonomy is still autonomy. The constraint is the point.
AI in Mathematics Is Forcing Big Questions
Terence Tao’s insight: formalization (proof assistants) removes the need to trust authority. You verify the proof step by step. That’s not just a math problem — it’s the architectural pattern for multi-agent collaboration with unknown contributors. When you can’t vouch for everyone in the system, verification-as-trust is the only way to scale beyond a small trusted group.
The counter worth sitting with comes from Maia Fraser: the struggle to understand builds human intuition you can’t shortcut. Optimize for output at the cost of the struggle and you build a system that’s dependent on what it can’t explain. The fastest path to an answer isn’t always the one that grows judgment. That applies to agents too.
🪨