Claude 4 API: Tool Use Streaming & Prompt Caching vs Cursor 1.0

“The primitive here is clean: incremental tool-call deltas over SSE, and a cache-control header you attach to prompt segments to pin them server-side. The DX bet is that complexity lives in the HTTP layer, not in a new SDK abstraction — you opt in per-request, no new mental model required. The moment of truth is calling `stream=true` on a tool-use request and watching partial JSON arguments arrive before the model finishes thinking, which actually matters for agent loops where you want to dispatch work early. This is not a weekend-script replacement — implementing correct incremental JSON parsing for partial tool arguments plus a reliable distributed cache with 2M token capacity is a real engineering problem Anthropic has solved for you. The specific decision that earns the ship: cache invalidation is explicit and cache hits are reflected in the usage object, so you can actually measure what you're saving instead of guessing.”

“The primitive here is a stateful, async coding agent that can hold context between your sessions and execute tasks in the background while you stay in flow — not a chatbot bolted onto a text editor. The DX bet is that memory and async execution should be editor-level primitives, not plugin afterthoughts, and that's the right call. First-10-minutes test: you open a project, the memory system picks up your conventions without a config file, and you can fire off a background task and come back to a diff. The weekend-script alternative collapses here — wiring persistent context, a sandboxed execution environment, and a real editor integration yourself is weeks of work, not a weekend. The specific decision that earns the ship is making background agent a first-class UI surface rather than a terminal command, which means it actually gets used.”

“Direct competitors are OpenAI's cached completions and Google's context caching in Gemini 1.5 — both shipping for months — so Anthropic is catching up, not leading. The specific scenario where this breaks: cache hit rates depend entirely on prompt structure, and developers who dynamically compose system prompts (inserting user-specific context at the top) will see near-zero cache utilization and pay full price while assuming they're saving money. The prediction: this feature doesn't get killed — it becomes table stakes infrastructure and Anthropic wins by having the largest cache window (2M vs. competitors' current limits). What would have to be true for me to be wrong: OpenAI ships a 10M token cache window before Anthropic's ecosystem matures, commoditizing the advantage. Still a ship because the streaming tool-use delta is genuinely differentiated — no competitor has clean partial-argument streaming for tool calls yet, and that changes agent loop architecture in ways that matter.”

“Direct competitors are GitHub Copilot Workspace, Windsurf, and Zed AI — Cursor's moat is the editor integration depth and the fact that they've been iterating in production with a large paying user base for over a year, not a demo environment. The scenario where this breaks is long-horizon background tasks on large polyglot monorepos: the agent context window fills, memory retrieval halts, and you get a half-applied diff with no clean rollback. That's not a theoretical failure mode, it's the current ceiling. What kills this in 12 months isn't a competitor — it's GitHub shipping a credible Copilot Workspace v2 with VS Code-native agent loops, which Microsoft has every distribution incentive to do. What would have to be true for me to be wrong: Anysphere ships a proprietary fine-tuned model that meaningfully outperforms the commodity frontier models they're currently wrapping, creating a performance moat that distribution alone can't replicate.”

“The thesis this bets on: by 2027, the dominant AI application architecture is a persistent agent with a large, stable context (tools, memory, instructions) that gets reused across thousands of user interactions — making context I/O cost the primary unit economics lever, not generation cost. The dependency that has to hold: agents don't collapse back to stateless chatbots, and context windows keep growing faster than per-token prices fall. The second-order effect nobody's talking about: prompt caching at 2M tokens makes it economically viable to give every enterprise user a fully-loaded, role-specific agent context at request time — which shifts competitive differentiation from 'who has the best model' to 'who has the best cached context corpus,' effectively making knowledge curation the new moat. This tool is riding the trend of context-window expansion-as-infrastructure, and it's on-time, not early — but the streaming tool-use primitive is ahead of the curve on agent loop efficiency. The future state where this is infrastructure: every production agentic system has a cache manifest the same way it has a CDN config.”

“The thesis is falsifiable: by 2027, the primary unit of software development is the task, not the keystroke, and developers manage fleets of async agents rather than writing code line by line. Background agent is the first editor-level implementation of that bet that's actually in production at scale, not a demo. What has to go right: agent reliability on real-world codebases has to improve from 'impressive demo' to 'trustworthy collaborator,' which requires both model capability gains and sandboxed execution that doesn't corrupt state. The second-order effect that matters isn't that developers get faster — it's that the ratio of senior-to-junior engineers a team needs shifts, because a senior can now supervise five parallel agent threads instead of writing code themselves. Cursor is riding the 'ambient compute replacing synchronous interaction' trend and they're on-time, not early — the infrastructure was ready, they just executed. The future state where this is infrastructure: every PR in a mid-size eng org has an agent trail attached, and code review becomes agent-output review.”

“The buyer is the engineering team at any company running Claude in production with long system prompts or multi-step agents — this comes out of the AI infrastructure budget, not a new budget line, which means no procurement friction. The pricing architecture is sound: cache reads at ~90% discount means the savings are real and measurable in the first billing cycle, which creates immediate retention — developers who restructure prompts to maximize cache hits are now architecturally coupled to Anthropic's caching implementation. The moat question is the honest one: this is infrastructure that OpenAI and Google will match, so the defensible position isn't the feature itself but the ecosystem of developers who've restructured their codebases around it. What survives a 10x model price drop: the streaming tool-use architecture, because that's about latency, not cost. The specific business decision that makes this viable is pricing cache reads as a separate SKU — it lets Anthropic capture value from high-volume production workloads without losing price-sensitive experimenters.”

“The buyer is an individual engineer or an engineering team lead pulling from a software tools budget — this is not a murky enterprise sale. Pricing architecture is clean: the free tier creates adoption, Pro at $20 captures the individual who hits the wall, and Business at $40 creates the team expansion motion with audit and admin controls. The moat question is the real one: right now they're wrapping Claude and GPT-4o, so the model isn't the moat — the moat is editor integration depth, the trained memory corpus attached to each user's codebase, and the switching cost of rebuilding your project memory elsewhere. That's real but fragile. What stress-tests the business: if Anthropic or OpenAI ships an IDE-native agent experience directly, Cursor's distribution advantage erodes fast. The specific decision that makes this viable is the memory layer — if that data becomes genuinely proprietary and personalized over time, they have a data flywheel that model providers can't replicate without the same surface area.”