Agent Swarms: The AI Architecture That Will Reshape Patent Work

This was a pivotal week in AI. Anthropic and OpenAI shipped major releases within minutes of each other. Cursor demonstrated autonomous agent swarms building production software. New benchmarks confirmed that AI systems can now sustain complex work for hours at a time.

If you were buried in case work and missed it, here's the rundown and why it matters.

Anthropic's Opus 4.6: Sixteen Agents Built a C Compiler

Anthropic released Claude Opus 4.6, which set the new state of the art on ARC-AGI, the hardest general reasoning benchmark in AI. It also has a new capability called "agent teams" — multiple Claude instances working in parallel on shared projects.

To demonstrate, Anthropic tasked Opus 4.6 with building a C compiler from scratch. The 16 instances coordinated across nearly 2,000 sessions over two weeks with minimal human oversight. The result after $20,000 of API spend: a 100,000-line Rust-based compiler that compiles the Linux kernel. 

The coordination pattern that mirrors litigation teams

One agent acts as team lead, decomposing problems, assigning tasks to specialists, and synthesizing outputs. The others work in parallel.

That maps directly to how patent litigation teams operate. A lead attorney defines strategy while associates handle parallel workflows: prior art searches, claim construction, prosecution history review.

GPT-5.3 Codex: The Model That Helped Build Itself

Minutes after Anthropic’s announcement, OpenAI released GPT-5.3 Codex, the first model that played an instrumental role in its own development. Engineers used earlier versions to debug, evaluate, and improve the production model. The result is 25% faster and uses fewer compute resources. 

Recursive self-improvement is no longer a thought experiment. It's a shipping methodology.

Your AI vendor evaluation has a shorter shelf life than you think

When AI systems contribute to their own improvement, development cycles compress. Capabilities that would take years arrive in months. For practitioners evaluating AI tools, the implication is straightforward: vendor evaluations and AI strategies that don't account for rapid capability growth risk being outdated before they're implemented.

Cursor's Agent Swarms: An Autonomous Engineering Org

Cursor recently published their research into scaling long-running autonomous coding agents. They built a new agent harness to orchestrate thousands of agents simultaneously.

To demonstrate in practice, roughly 2,000 concurrent agents operated autonomously for a full week, producing 3 million lines of code for a functioning web browser — including a custom rendering engine, CSS cascading logic, and a JavaScript VM. 

From single prompts to autonomous teams

Patent litigation involves large-scale analysis that currently requires a team working over weeks. Prior art searches across millions of patents. Claim charts across dozens of references. Agent swarms that coordinate specialized roles and sustain that coordination over extended periods are a fundamentally different tool than single-prompt AI assistants.

METR Benchmarks: AI Task Duration Is Growing

METR tracks time horizons, the length of tasks AI agents can autonomously complete. Between 2019 and 2025, those horizons doubled every seven months. Between 2024 and 2025, the rate compressed to every four months. Current models sustain autonomous work for over three hours.

The automation window for patent workflows just opened

The practical threshold for patent work is multi-hour sustained analysis. A comprehensive prior art search or invalidity review takes hours, not minutes. As AI task horizons cross that threshold, the range of workflows that can be meaningfully automated expands dramatically.

What Comes Next

Agent swarms that orchestrate complex tasks over days just went from research paper to production. Apply that to patent litigation and tasks like these become feasible:

• 1,000 patents. Every competitive product on the market with claim charts drafted, reviewed, and ranked within a week.
• Full due diligence on an acquisition target's IP portfolio. Validity, enforceability, competitive coverage in 48 hours.
• Every patent filed in a technology space over the last decade mapped, clustered, and scored for FTO risk overnight.

These aren't hypotheticals on a five-year roadmap. The coordination architecture, the sustained task duration, and the autonomous quality-checking patterns all shipped this week — in software engineering. The transition to legal is already underway, and the timeline is compressing.

At &AI, we're building the future of post-grant patent work for teams that refuse to wait for the industry to catch up. If that sounds like you, let’s talk.