Claude's Infinite Context and the Privacy Bill

Anthropic's "Code with Claude" developer conference yesterday was pitched as a showcase for the future of AI-assisted software engineering. Infinite context windows. Multi-agent orchestration. A "dreaming" capability where agents learn from past sessions. A new compute deal with SpaceX that's enabling doubled rate limits starting now.

The developer community is excited. I'm interested, and I have questions.

Let me be clear about sourcing up front: I'm working from Anthropic's own announcement page and a secondary breakdown from the WorldofAI channel, whose coverage is solid but is commentary, not a press release. Where specs are unconfirmed, I'll say so. There's enough here that's real to be worth unpacking without papering over the gaps.

What Anthropic Actually Announced

The confirmed news: Anthropic is doubling Claude Code's per-five-hour rate limits across all paid tiers — Pro, Max, Teams, and Enterprise — and removing the peak-hour throttling that's been frustrating Pro and Max users. API limits for Claude Opus are also getting a significant bump. Anthropic confirmed on its news page that this is enabled by a new compute partnership with SpaceX's Colossus 1 data center. The specific specs being cited — 300+ megawatts, 220,000+ Nvidia GPUs — are from the WorldofAI breakdown and secondary reports; Anthropic has not, as of this writing, confirmed those numbers publicly. The compute deal is real. The precise scale is still floating.

Beyond the rate limits, Anthropic announced several capabilities under the Claude Managed Agents umbrella: multi-agent orchestration (a lead agent coordinating parallel specialized agents), an outcome loop for iterative self-correction, webhook support for integrating agents into external tools and workflows, and the "dreaming" feature — agents reviewing prior sessions to improve future performance.

And then there are the three pillars Anthropic is flagging for its next generation of models: higher judgment and "code taste" (engineering quality beyond benchmark scores), infinite context windows, and deeper multi-agent coordination. Whether this lands in a Claude 5 family or something else entirely — the rumored "Mythos" system has been floating in speculation — Anthropic didn't commit to a timeline. Prediction markets have been pricing a Claude 5 release no earlier than late summer, but those numbers shift daily and I won't nail them to the wall.

The Architecture Question Nobody at the Conference Asked

Here's where I stop being a neutral relay and start being me.

When I read "agents can review past sessions and experiences to improve future outcomes," my first question isn't "cool, how fast will it code?" It's: what exactly is being retained, where, under what conditions, and who controls it?

The "dreaming" feature is genuinely interesting as a capability. An agent that learns from its mistakes across sessions — that evaluates what went wrong, adjusts its future behavior — is a meaningful step beyond a stateless chatbot. But that improvement loop requires something to loop over. Prior sessions. User interactions. Decisions made and outcomes evaluated. That's a data surface, and it's one that Anthropic has been quiet about in terms of specifics.

Similarly: multi-agent orchestration, where a lead agent coordinates parallel sub-agents working simultaneously across front-end, back-end, debugging, and research tasks. The attack surface there scales with the agent count. Each sub-agent is a trust boundary. Each one is calling tools, potentially accessing repositories, evaluating rubrics, passing outputs up to a coordinating layer. In a well-architected system, those handoffs are logged, permissioned, and auditable. The question is whether Anthropic is building those controls in from the start or treating them as something to retrofit once developers have built workflows on top.

I'm not saying the architecture is broken. I'm saying I don't know enough about it yet to tell you it isn't — and that's information you should have before you start piping your work projects through it.

What "Infinite Context" Actually Means for You

The phrase "infinite context window" is doing a lot of marketing work. What Anthropic is describing — at least based on what came out of the conference — is a model capable of retaining and reasoning across massive amounts of memory: repositories, projects, multiple sessions, tool outputs, all without the context resets that currently make agents frustrating for long-horizon work.

That's a real problem being solved. If you've ever used Claude Code for anything more than a single-session task, you know the amnesia problem. The agent forgets what it did three conversations ago. You re-explain your architecture. It re-makes decisions you already corrected. Fixing that would make the tool dramatically more useful.

But "infinite context" also means the model is holding more about you and your work for longer. That's a trade-off, not a free upgrade. For most personal use — someone using Claude to help draft a presentation or debug a personal project — the practical risk is low. For anyone running Claude through professional workflows that touch proprietary code, internal systems, or client data, it's a question worth asking your organization's security team before it becomes their problem to untangle.

As one early observer of the conference put it in Anthropic's own framing, the goal is for agents to work "across repositories, projects, sessions, and tools without constantly losing context." That's the pitch. The privacy policy is what tells you what happens to everything they're retaining to make that work.

The Infrastructure Bet

The SpaceX compute deal is the detail that makes the rest of this plausible. Anthropic has been capacity-constrained — the rate limits that frustrated developers weren't a product decision, they were a supply problem. Adding a major compute partner, whatever the exact specs turn out to be, changes what Anthropic can actually offer at scale.

What's notable is the breadth of what Anthropic disclosed alongside the SpaceX deal: compute agreements already in progress with Google, Amazon, Broadcom, Microsoft, Nvidia, and Fluid Stack. That's not a company hedging its bets — that's a company that was previously concentrated and is now deliberately distributing. Reading between the lines, the rate limit problems were at least partly a story about infrastructure dependency, and Anthropic is correcting for it.

The orbitable compute idea — SpaceX-hosted AI infrastructure in space — is the kind of announcement that exists to generate headlines and isn't worth treating as a product roadmap yet.

The Real Tension in the Room

Anthropic is, by its own public positioning, a safety-focused AI company. It's also, based on what came out of this conference, building toward autonomous software engineering systems that learn from your sessions, coordinate across parallel agents, self-correct against internal rubrics, and maintain persistent memory across your work.

Those things are not incompatible. But they do require specifics that Anthropic hasn't fully supplied yet: What data is retained for dreaming and for how long? What are the permission models for multi-agent tool access? What does the audit trail look like for an outcome loop that's iterating on your codebase autonomously? Where does human oversight re-enter a long-horizon agentic task?

The conference summary that came through WorldofAI's breakdown described the direction as moving Claude from "chatbot" to "fully autonomous software engineering system." That's accurate, and it represents a real category shift — not in the sci-fi sense, but in the practical sense that the trust model for a chatbot and the trust model for an autonomous agent with persistent memory are genuinely different. You can close a chat window. An agent with session memory that's "dreaming" over your prior work is a different kind of relationship.

The practical thing to do right now, if you're a Claude user who cares about this: read Anthropic's current privacy policy and usage terms before you expand into the new agent features. Check whether your organization has a policy on agentic AI tools. Ask, specifically, what "session memory" means in the context of whatever plan you're on. These questions have answers — or should — and getting them now is faster than untangling them after you've built a workflow around a tool you don't fully understand.

Anthropic is building something that could genuinely change how a lot of work gets done. That makes understanding what you're agreeing to more important, not less.

Rachel "Rach" Kovacs is Buzzrag's cybersecurity and privacy correspondent.