Code w/ Claude 2026 — Managed Agents Leap #
- Code w/ Claude 2026 — Live blog (Simon Willison, 2026-05-06) — Willison’s real-time notes from Anthropic’s developer conference. Key announcements: doubled rate limits for Pro/Max/Enterprise (peak-hours throttling dropped); Managed Agents Dreaming (background memory review, research preview); Outcomes (grader-agent evaluates against provided examples, public beta); Multiagent orchestration (up to 20 unique agent IDs in coordinator config); new Desktop app for Mac with iPhone/iPad integration.
- New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration (Anthropic, 2026-05-06) — Official technical detail. Dreaming: scheduled background process — after sessions end, a dreaming agent reviews interaction transcripts, extracts patterns, and curates memory stores without user input. Outcomes: provide 1–3 examples of “good” output; a grader agent evaluates every agent run against them. Multiagent: coordinator agent dispatches tasks to a fleet of specialised subagents; each agent ID maintains its own context and memory.
- Claude Code is getting higher usage limits, doubled for most users (9to5Google, 2026-05-06) — Rate limit doubling is the immediate developer-experience win. The 5-hour rolling window that was the primary Claude Code friction point for intensive sessions is significantly relaxed.
- Inside Anthropic’s 2026 Developer Conference (Every.to) — Conference narrative: Anthropic positioned the event as a transition from “Claude is a tool” to “Claude is an agent that can manage other agents.” Dreaming is the most structurally novel announcement — it implies agents that improve between sessions without human intervention.
- Using Claude Code: The Unreasonable Effectiveness of HTML (Simon Willison, 2026-05-08) — Willison links Thariq Shihipar’s (Anthropic Claude Code team) piece advocating HTML over Markdown as the default output format to request from Claude. Argument: HTML allows richer semantic structure, interactive elements, styled tables, and expandable sections that Markdown cannot represent. Willison notes he had defaulted to Markdown since GPT-4 days when token efficiency mattered — that constraint is now obsolete.
- Hacker News discussion (Hacker News) — Community response is strong: multiple practitioners note this is already a working pattern for dashboards, analysis reports, and code reviews. The key phrase: “ask Claude to make it rich, interactive, and clear” rather than specifying format constraints.
Cross-links #
- Vibe Coding Approaches Managed Agents Dreaming is the first Anthropic-native implementation of “agents that self-improve between sessions” — the agentic engineering discourse has been discussing this pattern theoretically; Dreaming is the concrete product instantiation.
- Claude Integrations Multiagent orchestration (20 agent IDs, fleet deployment) is the infrastructure backbone enabling the financial services workflow agents announced May 5 — same capability, different vertical packaging.
- Emerging theme: Dreaming is the most architecturally novel announcement of 2026 so far — an agent that reviews its own past interactions and improves its memory without a human asking it to. This is qualitatively different from user-configured CLAUDE.md. Worth tracking whether enterprise users adopt or resist autonomous memory evolution.
- Quality signal: Thariq Shihipar is on the Anthropic Claude Code team — the HTML > Markdown recommendation is based on direct model observation, not practitioner experimentation. Higher authority than community tips.
- Keyword suggestion:
"claude managed agents" dreaming outcomes — Anthropic-specific terminology for the new capability tier. - Gap: No coverage of Code Review, CI auto-fix, Security Reviews, Remote Agents, or Routines — the other features mentioned at the conference. Rate limits and Dreaming consumed the editorial attention.