Code with Claude 2026: multi-agent orchestration, routines and dreaming

At the Code with Claude conference on May 6, 2026 in San Francisco, Anthropic introduced the next iteration of its agent platform. Three terms carried the day: multi-agent orchestration, outcomes and dreaming. Alongside that, concrete Claude Code updates — routines, push notifications, a new TUI mode and a claude project purge command.

Multi-agent orchestration: fleets instead of single agents

Until now a Claude Managed Agent was a single worker with tool access. With multi-agent orchestration the model shifts: a task is split across several specialized sub-agents that run in parallel or in sequence, exchange intermediate results and converge on a consolidated outcome.

Anthropic’s demo example: a research-manager agent decomposes a market-analysis request, delegates data points to research sub-agents, has a writer sub-agent draft a first version, and a reviewer agent checks the draft against the outcome. What used to run as a one-shot prompt with a long context becomes a coordinated mini-team.

Outcomes — declarative instead of step-by-step

Outcomes is the second pillar. Instead of giving the agent step-by-step instructions, you define what the result should look like — for example: “Pitchbook on 12 slides, with updated Q1 2026 numbers, at least three comparison charts, style = corporate.template”. Claude then iterates on its own until that outcome is met, logging per iteration which criteria are still open.

In a coding context this isn’t radically new — test-driven iteration has existed for a while. In a business context (pitchbooks, reports, KYC checks) it’s the point where declarative outcomes start replacing classical prompt chains.

‘Dreaming’ — self-inspection of past sessions

The conference’s new term: dreaming. Managed Agents gain the ability to use their own completed sessions as reflection material. Between active tasks, the agent analyzes past solution paths, identifies recurring failure patterns, and adjusts its internal approach.

What’s technically behind this Anthropic hasn’t disclosed yet — likely a combination of session logs, an eval-pipeline pass, and a memory update. Conceptually it’s what’s discussed in agent research as a “self-improvement loop”, but Anthropic is the first major vendor to ship it as a production feature.

Claude Code updates from the May drop

In parallel to the conference, Anthropic shipped a substantial Claude Code update. The most important changes:

Routines. Repeatedly parameterizable workflows that bundle several CLI steps into a named flow. Related to the skill concept from the Anthropic API but anchored in the CLI world.

/tui and the tui setting. If you experienced flicker in the classic Claude Code terminal mode, you can now switch to a TUI renderer mid-session — same conversation state, but a flicker-free fullscreen interface.

Push notifications. With remote-control config enabled, Claude Code can send push notifications to a mobile device — for example when a long build or an agent run completes. “Push when Claude decides” is an additional threshold where the model itself decides when a notification makes sense.

claude project purge [path]. Deletes all project-related state — transcripts, tasks, file history, config entry. Useful for forks and clean re-runs where stale tasks would otherwise leak into the new context.

Bedrock service tier via env. ANTHROPIC_BEDROCK_SERVICE_TIER passes default, flex or priority through to AWS Bedrock — relevant for teams that use Claude via Bedrock instead of directly.

Reading between the lines

The pile-up of these three big concepts (orchestration, outcomes, dreaming) plus concrete CLI features is clearly designed to position Anthropic in the agent market — exactly while OpenAI courts the same audience with “Agent Mode” and Google with Gemini Agents. The open question: how much of this is actually production-ready? With multi-agent frameworks from 2024/2025 (LangGraph, CrewAI, AutoGen) it became clear that setup is easy, but the step into production breaks on eval, observability and cost control.

Anthropic addresses exactly that gap with outcomes (eval-by-spec) and dreaming (self-eval). Whether it holds up under real client work has to be shown over the coming weeks — the practical test with real mandate work is still ahead for many agencies.

What you can do now

If you use Claude Code locally: Update to the latest CLI, try claude /tui for flicker-free sessions and read the routines docs — your first workflows can often be consolidated immediately.

If you’re evaluating Managed Agents: Multi-agent orchestration and outcomes are worth a pilot use-case — preferably one whose result is clearly measurable (report generation, KYC check, code audit). I’d watch dreaming for now, not plan it as a production feature.

If you sell AI agents to clients: These terms become marketing standard from now on. For pitches it’s worth understanding outcomes as a concept — clients will ask about it.

Entdecke mehr

Blog

Claude Code: How to enable Fast Mode for Opus 4.6 — and what it costs

Fast Mode for Opus 4.6 is now extra-usage only. Here's how to enable it — and what it really costs you.

Glossar

Ensemble / Multi-Model Orchestration

Ensemble means combining several deliberately varied LLM runs or models whose findings complement each other. Multi-model orchestration drives these runs via orchestrators with sub-agents, so the union of results is larger than any single run.

Lexikon

Function Calling / Tool Use

How an LLM uses tools: define a tool as a schema, the model picks the function and arguments, the result returns to the chat — the basis of every agent.

Themenuebersicht

KI-WerkzeugeCoding-CLIKI-KonzepteAgenten