Anthropic held their second annual developer conference, Code with Claude, on May 6, 2026 at Pier 48 in San Francisco. No new model was announced — the theme was entirely about developer velocity and making the existing models work better in production. API volume on the Anthropic platform is up 17x year-over-year.

Like I did for Google Cloud Next, I let an AI summarize all the sessions so you can scan what happened and pick what’s worth watching. All session recordings are now on the Code with Claude 2026 YouTube playlist on the @claude channel.

🤖 Code with Claude 2026 — The Developer-First Conference

Conference: Code with Claude 2026
Date: May 6, 2026
Location: Pier 48, San Francisco, CA
Theme: “Making our products work better for you” — from prototype to production, at scale
Follow-on events: London (May 19) · Tokyo (June 10)

All recordings: YouTube playlist · @claude on YouTube

The Big Picture

Anthropic opened with a clear signal: this wasn’t a model launch day. Chief Product Officer Ami Vora framed the entire conference around developer outcomes — teams shipping 10x faster, execs getting their hands dirty with code again, agents running for hours instead of minutes. The stat that set the tone: API traffic is up 17x year-over-year.

The major news on the day fell into three buckets. First, a massive compute expansion: a partnership with SpaceX’s Colossus 1 data center in Memphis, unlocking 300+ megawatts and 220,000+ NVIDIA GPUs — which directly funded doubling Claude Code’s rate limits across Pro, Max, Team, and Enterprise. Second, three new capabilities for Claude Managed Agents: Dreaming (a scheduled process that reviews past sessions and self-improves), Outcomes (set success criteria and let Claude iterate), and Multiagent Orchestration (fleets of specialized agents on a shared filesystem). Third, a slate of Claude Code updates centered on async, automated workflows: Routines (scheduled prompts that wake up and file PRs), CI Auto-fix, Code Review, Remote Agents, and Security Reviews.

The underlying message across all sessions: the era of “minutes of agent time” is over. Developers are now running agents for hours — or overnight — and the products are being redesigned around that reality.

Opening Keynote

📺 Watch on YouTube · Session page

Speakers: Ami Vora (CPO), Dianne Na Penn (Head of Product, Research), Angela Jiang (Head of Product, Claude Platform), Katelyn Lesse (Head of Engineering, Claude Platform), Cat Wu (Head of Product, Claude Code), Boris Cherny (Head of Claude Code)

The keynote opened with Ami Vora sharing developer velocity stories — Scott MacVicar’s team at Stripe, Felicia Curcuru’s team at Binti — then pivoted to a research highlight from Dianne Na Penn: Anthropic’s Mythos model reading the OpenBSD source tree and finding a 27-year-old vulnerability. Dianne framed the model maturation story: reliable tool use, long context, computer use, adaptive thinking, visual design, and agentic loops have all moved from aspirational to real. Opus 4.7 has “a real taste for visual design.” The classic advice from the research team: design for the next model — build things that don’t quite work today on the assumption that model upgrades will make them work.

Katelyn Lesse and Angela Jiang introduced the Claude Platform section, covering the Advisor strategy — Sonnet calling Opus on demand as a smarter advisor, delivering frontier model quality at 5x lower cost for one early customer. Then the three new Managed Agents features were announced: Multiagent Orchestration (lead agent + specialist subagents working in parallel, shared filesystem), Outcomes (define success criteria and let Claude iterate), and Dreaming (overnight session review + memory curation — demo showed a descent-playbook.md file created from a lunar landing drone simulation). Multiagent Orchestration and Outcomes entered public beta; Dreaming is a research preview requiring access request.

Cat Wu took the Claude Code section, boasting about production deployments at Shopify and Mercado Libre (23,000 engineers, targeting 90% autonomous coding by Q3 2026). Cat walked through the four Claude Code surfaces — CLI, IDE, Desktop app, and the web — plus new features: Code Review (used by every team at Anthropic), Remote Agents (control your laptop from your phone), CI Auto-fix, and Security Reviews. Boris Cherny closed with a live demo of the Desktop app running multiple async sessions simultaneously, showing Claude discovering and fixing an edge-case bug in a refund feature before a human would have caught it. He introduced Routines — higher-order prompts that run on a schedule, on API trigger, or in response to GitHub events — framing them as the foundation for async engineering: “wake up to PRs ready to merge.”

What’s New in Claude Code

📺 Watch on YouTube · Session page

Speaker: Dickson Tsai (Claude Code team)

A focused twenty-minute summary of everything that shipped in Claude Code over the past year: what was built, why, and how to get started. Covers the full arc from CLI to IDE to Desktop app to web, with the Claude Agent SDK as the shared foundation external developers can build on. Key shipping highlights: Code Review, Remote Agents, CI Auto-fix, Security Reviews, and the Routines system. The session is the right starting point if you want a structured tour of the full feature set rather than the broader keynote framing. Docs and release notes for everything covered are in the What’s New log.

Caching, Harnesses, and Advisors: Building on Claude at GitHub Scale

📺 Watch on YouTube · Session page

Speakers: Mario Rodriguez (CPO, GitHub), Brad Abrams (Anthropic)

GitHub’s Copilot team ships Claude to millions of developers across chat, CLI, coding agent, and code review — one of the most demanding operator deployments on the Anthropic platform. Mario Rodriguez and Brad Abrams broke down how GitHub pushes quality up and costs down at massive scale. Three techniques took center stage.

Prompt caching is fundamental at this scale: structuring prompts so that expensive prefix computations can be reused across requests dramatically cuts cost without touching model quality. Harness tuning — the scaffold connecting the model to tools, environment, and memory — is often more impactful than fine-tuning the model itself; optimizing the harness alone can push a model from 30th to 5th on an engineering benchmark. The Advisor strategy is the newest of the three: routing selected requests to Opus as an advisor for Sonnet — effectively giving Sonnet access to frontier-quality judgment for the specific moments where it matters, while keeping the average cost per request low. The session is a practical playbook for any team running Claude at scale where cost and quality both matter.

How to Get to Production Faster with Claude Managed Agents

📺 Watch on YouTube · Session page

Speakers: Jess Yan, Lance Martin (Anthropic)

A deep-dive into Claude Managed Agents — Anthropic’s platform for teams who want to ship agents “10x faster” by getting best practices (memory, orchestration, persistence) out of the box rather than building them from scratch. The session walked through all three of the day’s new features with implementation context.

Multiagent Orchestration (public beta): design a lead agent that delegates to specialist subagents running in parallel on a shared filesystem — shown with a Commander + Detector + Navigator pattern for a lunar drone landing simulation. Outcomes (public beta): instead of writing a fixed task prompt, you define what success looks like and let Claude iterate autonomously. Dreaming (research preview): a scheduled overnight process that reviews the agent’s past sessions, extracts patterns across them — recurring mistakes, workflows the agent converges on, preferences shared across a team — and curates new memories. The demo created a descent-playbook.md artifact from session history. The session also covered webhooks and the full Managed Agents documentation. Request access to Dreaming.

The Expanding Toolkit

📺 Full playlist · Session page

Over the last year, capabilities that previously required heavy scaffolding have moved into the model itself: reliable tool use, context management, writing and running code, computer use. This session surveyed that shift and its practical implications for how developers should architect their Claude-based systems. The key insight: many patterns that were workarounds a year ago are now unnecessary — you can lean on the model to handle things directly that previously required explicit scaffolding. Best practices for tool use, context management, and agentic loop design with current models. Pairs well with Anthropic’s best practices documentation for Claude Code.

Live Coding with Bun and Claude Code

📺 Watch on YouTube · Session page

A live coding session on the main stage pairing the Bun JavaScript runtime with Claude Code. Bun’s speed and bundler-inclusive design make it a natural fit for the kind of rapid, iterative development Claude Code enables. The session was a practical demo of the full loop: starting from a prompt, generating and running code, iterating on failures, and shipping a working result — using Bun’s fast feedback cycle to keep the agent loop tight. A good watch for JavaScript/TypeScript developers looking to see Claude Code in its native workflow.

Key Takeaways for Developers

1. Agents run overnight now — design for it. Boris Cherny’s framing: “a lot of code is going to be written in an async way.” Routines, CI Auto-fix, and Dreaming are all built around the assumption that your agent is working while you sleep, and filing PRs or updated memories for you to review in the morning.

2. Routines replace the cron job pattern. Claude Code Routines let you save a prompt + repo + connectors as a single configuration and trigger it on a schedule, API call, or GitHub event. Common patterns: nightly issue triage, weekly docs drift detection, deploy verification. Docs · Blog

3. The Advisor strategy is a real cost/quality lever. Having Sonnet call Opus on demand — for the specific moments where frontier judgment matters — gives you frontier-quality outputs at a fraction of the cost of using Opus everywhere. GitHub validated this at scale.

4. Managed Agents abstracts the hard parts. Memory, orchestration, persistence, and multi-agent coordination are all bundled. If you’re building a long-running agent and doing this yourself, Managed Agents is worth evaluating — it’s Anthropic’s distilled best practices as a platform. Docs

5. Dreaming is the most novel idea from the day. The concept of a scheduled process that reviews an agent’s past sessions and curates memories — self-improvement without retraining — is architecturally interesting and practically useful for any agent that runs repeatedly on similar tasks. Still research preview, but worth watching.

6. The Claude Agent SDK is available to you. The same SDK powering the Claude Code IDE and Desktop app is available for external developers to build on. If you’re building a coding agent or developer tool, this is the foundation Anthropic ships on. Docs

7. Compute got cheaper overnight. The SpaceX Colossus deal isn’t just PR — it directly funded doubled rate limits for all paid Claude Code tiers. The peak-hours reduction on Pro and Max is gone. If you’ve been rate-limited, try again.

Announcements at a Glance

Sources

• Code with Claude — official event page
• Opening Keynote — YouTube (@claude)
• Code with Claude 2026 — full recording playlist
• Code w/ Claude SF 2026 recap — Anthropic blog
• Higher usage limits + SpaceX deal — Anthropic
• Introducing Routines in Claude Code — Anthropic
• Managed Agents docs
• Dreaming docs
• Simon Willison’s live blog
• VentureBeat — Dreaming announcement

Related

• [[Projects/Code with Claude 2026]]
• [[Topics/AI]]
• [[2026-05-06]]