The Agents Are Dreaming
On May 6, 2026, Anthropic rolled out what might be its most significant Managed Agents update to date — a feature set that fundamentally changes how autonomous agents operate in the enterprise. The headline feature is “Dreaming”: a scheduled background process that gives Managed Agents the ability to reflect on, iterate through, and eventually resolve complex problems without requiring constant human supervision or real-time interaction.
The term “dreaming” is deliberately evocative. Just as the human brain consolidates memories and processes information during sleep, Anthropic’s dreaming mechanism allows agents to operate on tasks during scheduled windows — working through multi-step reasoning chains, backtracking from dead ends, and converging on solutions in a structured, auditable loop. It’s a scheduled memory and computation process that runs independently of user-facing interactions.
“Dreaming turns managed agents from reactive tools into proactive problem-solvers,” the company stated in its announcement. “Instead of waiting for a prompt, agents can now actively work through assigned objectives during designated compute windows.”
How Dreaming Works
Behind the scenes, dreaming operates on a scheduled compute budget. Administrators configure when and for how long an agent can “dream” — typically during off-peak hours or between human work shifts. During these windows, the agent:
- Reviews its task queue and prioritizes unresolved objectives
- Runs exploratory reasoning chains across multiple paths in parallel
- Evaluates intermediate results against success criteria
- Commits solutions or surfaces blockers that require human input
The process is fully transparent: each dreaming session produces a detailed log of reasoning steps, dead ends explored, and decisions made. This audit trail is critical for enterprise compliance requirements — a common barrier to autonomous agent deployment in regulated industries.
Critically, dreaming is not just “thinking longer.” It’s a structured, resource-bounded process with configurable limits on compute time, token usage, and exploration depth. Organizations can set strict budgets per session, ensuring that dreaming never spirals into uncontrolled resource consumption. For more on Anthropic’s agent platform, see our complete guide to AI agents.
Outcomes-Based Evaluation: Measuring What Matters
Alongside dreaming, Anthropic introduced Outcomes-Based Evaluation (OBE) — a new framework for assessing agent performance that shifts the metric from “did the agent follow the right steps?” to “did the agent achieve the right result?”
Traditional agent evaluation has relied on step-by-step verification: checking that each tool call, API request, or reasoning step matches an expected sequence. OBE flips this model. It defines success criteria at the outcome level — “customer ticket resolved with satisfaction score > 4/5” or “quarterly report generated with all required sections” — and measures whether the agent hits those targets, regardless of the path taken.
This is a paradigm shift for enterprise agent deployments. It enables:
- Fuzzy success criteria — agents can take different approaches to the same problem
- Graceful degradation — partial success is measured and reported
- Autonomous optimization — agents can discover better strategies than the ones humans would prescribe
- Simpler evaluation pipelines — no need to instrument every intermediate step
OBE integrates directly with the dreaming system: after a dreaming session, the agent self-evaluates against its outcome criteria and reports a success/failure grade with supporting evidence.
Multi-Agent Orchestration in Public Beta
The third major piece of the announcement is the public beta of multi-agent orchestration — the ability to deploy multiple Managed Agents that coordinate on shared objectives.
Anthropic’s orchestration model uses a supervisor-agent architecture. For more on Anthropic’s expanding agent platform, see our complete guide to AI agents and the coverage of the FIS financial crime agent.
# Example configuration concept
orchestration:
supervisor:
model: claude-opus-4-sonnet
role: "Coordinate and delegate sub-tasks"
workers:
- agent: "data-collector"
model: claude-sonnet-4-5
tools: [web-search, api-query, database-read]
- agent: "analyzer"
model: claude-opus-4-sonnet
tools: [python-exec, chart-generation, report-writer]
- agent: "reviewer"
model: claude-sonnet-4-5
tools: [diff-check, compliance-scan, approval-request]
The supervisor agent decomposes high-level objectives into sub-tasks, assigns them to specialized worker agents, aggregates results, handles conflicts, and escalates exceptions. Worker agents can themselves leverage dreaming for their sub-tasks, creating a nested hierarchy of autonomous problem-solving.
This architecture mirrors the team structures already common in enterprise software development — and that’s exactly the point. Anthropic is positioning Managed Agents as collaborative team members rather than isolated tools.
Enterprise Implications
Taken together, these three features represent a significant maturation of the enterprise agent platform category:
| Capability | What It Enables |
|---|---|
| Dreaming | Off-hours autonomous problem-solving, reduced need for synchronous human interaction |
| Outcomes-Based Evaluation | Simplified QA, agent autonomy within guardrails, measurable ROI |
| Multi-Agent Orchestration | Complex workflows, specialization, human-like team structures |
For CIOs and engineering leaders evaluating agent platforms, the message is clear: Anthropic is betting that enterprise adoption requires agents that can work asynchronously, autonomously, and collaboratively — not just respond to prompts faster.
The timing is notable. With Google building “Remy” and OpenClaw already demonstrating persistent, always-on agents, the competitive landscape for the agent operating system is heating up fast. Anthropic’s Managed Agents update places a clear bet on the enterprise segment — where compliance, auditability, and structured workflows matter most.
Availability and Pricing
Dreaming, OBE, and multi-agent orchestration are available now in public beta for Managed Agents customers. Pricing varies by compute budget allocated to dreaming sessions, with tiered plans starting at $2,000/month for teams. Enterprise plans with custom SLAs, dedicated compute pools, and on-premises deployment options are also available.
Sources: The New Stack — Anthropic will let its managed agents dream, Anthropic Blog, and internal analysis.