What are the new Code with Claude London agent features? The Code with Claude London agent features are three new managed agent capabilities Anthropic announced at its developer conference on May 19, 2026: Multiagent Orchestration, which lets a lead agent dispatch parallel sub-agents across complex tasks, Outcomes, which lets developers define success criteria so agents can self-correct against a rubric, and Dreaming, which allows Claude to consolidate memory across past sessions. For commercial real estate (CRE) firms, these are not abstract developer features, they are the first time Anthropic has shipped production-grade primitives for the kind of long-horizon, multi-step, memory-dependent workflows that actually match how acquisitions, asset management, and portfolio operations run in practice. For broader context on agentic AI in CRE, see our pillar guide on AI tools for real estate investors.
Key Takeaways
- Anthropic announced three new managed agent features at Code with Claude London on May 19, 2026: Multiagent Orchestration, Outcomes (public beta), and Dreaming.
- Outcomes reports +8.4% task success on docx and +10.1% on pptx in Anthropic internal benchmarks via rubric-based self-correction by a separate grader model.
- Multiagent Orchestration enables fleets of parallel sub-agents, matching the natural structure of CRE workflows like multi-property due diligence and portfolio-wide compliance review.
- Dreaming consolidates past session transcripts into a reorganized memory store without modifying the original, enabling persistent client and asset memory across an asset manager's workflow.
- Anthropic also doubled Claude Code five-hour rate limits across Pro, Max, Team, and seat-based Enterprise plans and announced a 300+ megawatt SpaceX Colossus 1 compute partnership.
What Anthropic Shipped at Code with Claude London
Code with Claude 2026 ran across three cities: San Francisco on May 6, London on May 19, and Tokyo on June 10. The London keynote was the first physical European meeting for the developer community and featured live demos with the Anthropic team across three parallel programming tracks: Research, Claude Platform, and Claude Code. The headline announcements were the three managed agent features, alongside doubled Claude Code rate limits, the SpaceX Colossus 1 compute partnership bringing 220,000+ NVIDIA GPUs online within a month, and partner sessions from GitHub, Amazon, Datadog, Vercel, Cursor, and Microsoft.
Multiagent Orchestration: Parallel Sub-Agents for CRE Workflows
Multiagent Orchestration is the feature with the most immediate CRE impact. The pattern is simple: a lead agent decomposes a complex task and dispatches sub-agents to work in parallel, each with focused context and tool access. Anthropic's example was incident response, with a lead agent running an investigation while sub-agents fan out simultaneously through deploy history, error logs, metrics, and support tickets.
The CRE translation is direct. Consider a 50-property multifamily portfolio due diligence assignment. Today, that work is run sequentially: one analyst pulls T12 financials, then rent rolls, then lease abstracts, then property condition reports, then market comps. With Multiagent Orchestration, a lead agent can dispatch five sub-agents in parallel, each focused on one workstream across all 50 properties, with the lead agent reconciling and producing a consolidated investment memo. The same pattern applies to portfolio-wide loan covenant compliance, ESG reporting, insurance renewal benchmarking, and quarterly variance analysis. For more detail on AI-driven due diligence patterns, see our guide on AI real estate due diligence.
Outcomes: Self-Correcting Agents Against a Rubric
Outcomes is in public beta and uses a rubric-based separate grader model for self-correction. The developer defines what success looks like (correctness criteria, format requirements, tone, completeness), and the agent iterates until the grader signs off. Anthropic reported +8.4% task success on docx generation and +10.1% on pptx generation in internal benchmarks using this loop.
For CRE firms, this is the feature that finally takes a class of agent failures off the table. Today, an agent generating an investment committee memo will produce plausible-looking output that misses the specific metrics the committee expects (DSCR at year-3 stabilization, IRR net of fees and promote, cash-on-cash in year 1). With Outcomes, the firm can codify the committee rubric once and have the agent grade itself before delivering. The same applies to LP quarterly reports, loan submission packages, and broker BOV responses. Pricing the time saved is straightforward: every memo that does not have to be reworked by a principal is 30 to 90 minutes of senior-level time recovered.
Dreaming: Persistent Memory Across CRE Asset Lifecycles
Dreaming is the most experimental of the three features. The agent reads an existing memory store alongside past session transcripts, then produces a new, reorganized memory store with duplicates merged and new insights surfaced. The original input store is never modified, so a human can review the consolidation before adopting it.
The CRE workflow this unlocks is asset-level institutional memory. A CRE asset manager today operates across hundreds of sessions per asset over a five-to-seven year hold: acquisition, financing, leasing, capital projects, refinancing, disposition. Crucial context lives in one analyst's head, one Outlook folder, one Slack channel, and one shared drive. Dreaming gives an asset-management agent a path to build a consolidated memory store per asset that survives staff turnover, gets refreshed monthly, and is auditable because the underlying transcripts are never overwritten. For more on how persistent context reshapes portfolio operations, see our guide on AI multifamily underwriting.
Why CRE Should Care About the Compute and Rate-Limit Changes
Underneath the agent feature announcements, Anthropic doubled Claude Code's five-hour rate limits across Pro, Max, Team, and seat-based Enterprise plans and removed the peak-hours limit reduction on Pro and Max accounts. The infrastructure backstop is a 300+ megawatt compute partnership with SpaceX at the Colossus 1 facility, bringing 220,000+ NVIDIA GPUs online within a month.
For CRE firms running real workloads, rate limits are usually the silent ceiling on agentic deployments. A CRE acquisitions team that wanted to run Claude across every deal in the funnel would routinely hit the prior five-hour ceiling by mid-day. Doubled limits, combined with a new compute backstop, means agentic workflows that were previously throttled can now scale across an entire team. The CRE in-house implementation cost has just shifted from rate-limit engineering to genuine workflow design. If you're ready to transform your underwriting process with AI, The AI Consulting Network specializes in exactly this. Coverage of the broader CRE AI shift is available from JLL.
Three Implementation Patterns for CRE Firms This Quarter
- Pattern 1, Multiagent Due Diligence: Build a lead agent that takes a deal package as input and dispatches sub-agents for financial review, lease review, market comps, environmental and zoning, and capital needs. Reconcile the outputs into a structured investment memo. Best target: any acquisitions team running more than 5 deals per month.
- Pattern 2, Outcomes-Graded LP Reporting: Codify the LP report rubric (required sections, required metrics, tone, page count, attachment list). Run Outcomes against the rubric and only deliver when the grader signs off. Best target: any GP managing 3+ active funds with quarterly reporting cadence.
- Pattern 3, Dreaming for Asset Memory: Run monthly memory consolidation on each major asset's session history. The output becomes the briefing document for any new staff member, any quarterly review, and any disposition decision. Best target: any owner-operator with 50+ assets and analyst turnover.
CRE investors looking for hands-on AI implementation support can reach out to Avi Hacker, J.D. at The AI Consulting Network to design and deploy these patterns. The market context underscores the urgency: AI in real estate is projected at $1.3 trillion by 2030 at 33.9% CAGR per industry research, 92% of corporate occupiers have initiated AI programs, but only 5% report achieving most of their AI program goals. The firms moving from pilots to production agentic systems in 2026 are the ones who will capture the 7.2x leader-laggard gap. For a comparison of the underlying models powering these agents, see our AI model comparison CRE guide.
Frequently Asked Questions
Q: When are Multiagent Orchestration, Outcomes, and Dreaming available to Claude developers?
A: Outcomes is in public beta as of the May 19, 2026 announcement. Multiagent Orchestration and Dreaming were demonstrated at Code with Claude London with broader availability rolling out through the Claude Developer Platform. Verify exact tier and beta-flag access against Anthropic's current developer documentation before designing production workflows.
Q: Do these features require Claude Opus 4.7 or do they work with Claude Haiku 4.5?
A: Anthropic positioned all three features as managed agent primitives on the Claude Developer Platform rather than as model-specific capabilities. The underlying model choice typically matters more for reasoning depth than for orchestration mechanics. For CRE use, route long-horizon decision tasks to Claude Opus 4.7 and run high-volume sub-agent tasks on Claude Haiku 4.5 to control cost.
Q: How does Outcomes differ from prompt engineering and review loops we already run?
A: Outcomes runs a separate grader model against a developer-defined rubric and feeds back to the original agent for iteration, automated rather than human-mediated. The +8.4% docx and +10.1% pptx improvements Anthropic cited come from this self-correction loop running before output is delivered, not from a better single-pass prompt.
Q: Can Dreaming replace a CRE firm's CRM or asset management system?
A: No. Dreaming consolidates session-level conversational memory, it is not a system of record for rent rolls, leases, or financials. Treat Dreaming as a context layer that makes an asset-management agent smarter across hundreds of past interactions, with your existing CRE platform (Yardi, RealPage, MRI, AppFolio, Juniper Square) remaining the authoritative data store.
Q: What is the right way to start with Multiagent Orchestration in a small CRE shop?
A: Start with one workflow that has natural parallel structure, due diligence is the cleanest fit, and deploy a lead agent plus 3 to 5 focused sub-agents for that single workflow. Resist the temptation to build a generalized orchestrator on day one. Measure cycle time and quality before and after, then expand pattern by pattern. The 80/20 win is one well-designed orchestration, not a fleet of half-baked ones.