Anthropic on Tuesday unveiled a set of updates to its Claude Managed Brokers platform at its second annual Code with Claude developer convention in San Francisco, introducing a brand new functionality referred to as "dreaming" that lets AI brokers study from their very own previous periods and enhance over time — a step towards the form of self-correcting, self-improving AI methods that enterprises have demanded earlier than trusting brokers with manufacturing workloads.
The corporate additionally moved two beforehand experimental options — outcomes and multi-agent orchestration — from analysis preview into public beta, making them broadly accessible to builders constructing on the Claude platform. Collectively, the three options tackle what Anthropic says are the toughest issues in working AI brokers at scale: retaining them correct, serving to them study, and stopping them from changing into bottlenecks on advanced, multi-step work.
Early adopters are already reporting vital outcomes. Authorized AI firm Harvey noticed process completion charges improve roughly 6x after implementing dreaming. Medical doc evaluate firm Wisedocs lower its doc evaluate time by 50% utilizing outcomes. And Netflix is now processing logs from lots of of builds concurrently utilizing multi-agent orchestration.
The bulletins come at a second of extraordinary momentum for Anthropic. CEO Dario Amodei disclosed throughout a hearth chat on the convention that the corporate's development has outpaced even its personal aggressive inside projections.
Within the first quarter of 2026, Anthropic noticed what Amodei described as 80x annualized development in income and utilization — far exceeding the 10x annual development the corporate had deliberate for. API quantity on the Claude platform is up almost 70x 12 months over 12 months, and the common developer utilizing Claude Code now spends 20 hours per week working with the device.
"We tried to plan very nicely for a world of 10x development per 12 months," Amodei stated. "And but we noticed 80x. And so that’s the reason we’ve got had difficulties with compute."
How Anthropic's dreaming characteristic teaches AI brokers to study from their very own historical past
Dreaming is essentially the most novel of the three options and the one Anthropic is most keen to tell apart from typical reminiscence methods. Whereas the corporate launched agent reminiscence earlier this 12 months — permitting Claude to retain preferences and context inside and throughout particular person periods — dreaming works at the next stage of abstraction. It’s a scheduled course of that evaluations an agent's previous periods and reminiscence shops, extracts patterns throughout them, and curates these recollections so brokers enhance over time. It surfaces insights that no single agent session might see by itself: recurring errors, workflows that a number of brokers converge on independently, and preferences shared throughout a crew of brokers.
Alex Albert, who leads analysis product administration at Anthropic, defined the idea in an interview on the convention. He described dreaming as analogous to how folks inside organizations create abilities after working via a process. "They may do a workflow with Claude, and on the finish of that workflow, after they've iterated and zigzagged a bit bit, they wish to document that path from A to B," Albert stated. "A really related factor is occurring with dreaming — as a substitute of you manually creating the ability out of your expertise working with Claude, the mannequin is doing it, so it has that very same context for a future session."
Crucially, dreaming doesn’t modify the underlying mannequin weights. "We're not altering the mannequin itself via dreaming — it's not doing updates to the weights or something like that," Albert stated. As an alternative, the agent writes learnings as plain-text notes and structured "playbooks" that future periods can reference, making the whole course of observable and auditable by people. When requested in regards to the belief implications of brokers consolidating their very own information, Albert acknowledged that "there’s a stage of belief that you could place" however famous that each one recollections are inspectable and that smarter fashions are getting progressively higher at managing this course of. "They're studying to write down higher notes for his or her future self," he stated.
A reside demo confirmed AI brokers enhancing in a single day with out human steering
In the course of the keynote, the Anthropic crew demonstrated all three options reside on stage utilizing a fictional aerospace startup referred to as "Lumara" that wanted to autonomously land drones on the moon for useful resource mining. The crew configured a multi-agent system with three specialists — a commander agent chargeable for general mission success, a detector agent that recognized high-quality touchdown websites, and a navigator agent that dealt with protected drone flight and touchdown — and outlined successful rubric requiring delicate landings, clear floor, and sufficient gas reserves for a return journey to Earth.
An preliminary simulation throughout six hypothetical touchdown websites produced sturdy however imperfect outcomes. To enhance, the presenters triggered a dreaming session instantly from the Claude Developer Console. In a single day, the dreaming agent reviewed all previous simulation periods and wrote an in depth descent playbook — a complete set of heuristics drawn from patterns throughout a number of mission runs. When the crew ran a brand new simulation the next morning with the dreaming-derived playbook in reminiscence, the outcomes improved meaningfully on the websites that had beforehand underperformed.
"All we needed to do was simply have Caitlin press a button," stated Angela Jiang, Head of Product for the Claude Platform, referring to her colleague on stage. "All dreaming."
The demo illustrated how the three options compose collectively in observe. Multi-agent orchestration break up the advanced process throughout specialists with impartial context home windows. Outcomes offered the rubric in opposition to which a separate grader agent evaluated every run. And dreaming extracted classes throughout these runs to enhance future efficiency — forming what Anthropic describes as a steady enchancment loop that requires no human intervention between iterations.
Why Anthropic constructed a separate 'grader' agent to examine Claude's personal work
The outcomes characteristic, now in public beta, offers builders a method to outline what success appears like utilizing a rubric — a structural framework, a presentation customary, a model voice, or every other set of standards — after which lets the agent iterate towards that customary autonomously. What makes outcomes architecturally distinctive is its separation of considerations. When an agent completes its work, a separate grader agent evaluates the output in opposition to the developer-defined rubric in its personal impartial context window. As a result of the grader operates in a contemporary context, it’s not influenced by the working agent's reasoning or collected biases from the session.
When the grader identifies gaps between the output and the rubric, it pinpoints particularly what wants to vary, and the working agent takes one other go. This loop continues till the rubric standards are met — with out a human needing to evaluate every try.
Albert described Anthropic's broader verification technique as using "extra check time compute, extra fashions enthusiastic about an issue for longer, to examine over the work of one other." He acknowledged that having a mannequin examine its personal work raises cheap questions, however stated a contemporary context window reviewing accomplished work constantly outperforms asking the identical long-running thread to establish its personal bugs. "You’re going to get increased success for those who give that output to a contemporary Claude and say, 'what bugs do you see?'" he stated. "There’s nonetheless one thing to the eye" that degrades over very lengthy periods — a limitation he stated Anthropic is actively working to repair in future fashions.
The method mirrors methods already in use at GitHub. Mario Rodriguez, Chief Product Officer at GitHub, described throughout a separate speak on the convention how Copilot makes use of the same advisor sample with Claude fashions — pairing a smaller, cheaper mannequin as an executor with a bigger mannequin as a mentor. When the smaller mannequin encounters an issue past its functionality, it calls the bigger mannequin for steering, then continues executing by itself. Rodriguez stated the method delivers near-Opus-level intelligence at considerably decrease value, and that GitHub inserts critique fashions at three particular factors within the coding workflow: after drafting a plan, after a fancy implementation, and after writing assessments however earlier than working them.
Parallel AI brokers can now deal with duties too advanced for a single mannequin thread
Multi-agent orchestration, the third characteristic shifting to public beta, permits a lead agent to decompose a big process into subtasks and delegate every one to a specialist agent — every with its personal mannequin, system immediate, instruments, and impartial context window. Each step within the course of is traceable within the Claude Console, displaying which agent did what, in what order, and why.
The design offers every sub-agent an remoted context, which Anthropic says produces higher outcomes than having a single agent try to carry all of the complexity in a single thread. "Every sub-agent has its personal impartial thread and context window," the keynote presenters defined. "That is very intentional — we discovered that by splitting the work after which merging the outcomes, we get higher outcomes."
Albert supplied his personal heuristic for when multi-agent architectures make sense versus sticking with a single thread. "Parallel brokers are higher for investigation," he stated — conditions the place there may be numerous context that can finally be discarded. "In the event you're making an attempt to reply a selected query, you don't want all of the search outcomes from the areas the place it didn't discover the reply. You simply want the reply." He described spinning up disposable sub-agents for particular retrieval duties and bringing solely the end result again to the principle thread. More and more, he stated, the mannequin itself will determine when to parallelize. "Sooner or later, you received't actually care if it's one agent or multi-agent or no matter's occurring. You simply have a Claude that you simply're speaking to, and it’ll deploy the fitting structure mechanically."
Anthropic's greater wager: closing the hole between AI capabilities and real-world adoption
The three options arrive as a part of a broader platform push that Anthropic framed all through the convention as closing "the hole between what AI can do and what it's truly doing for folks." Ami Vora, Anthropic's Chief Product Officer, set the theme in her opening keynote, noting that whereas mannequin capabilities are advancing on an exponential curve, most organizations are nonetheless adopting AI on a linear path.
Dianne Penn, who leads product for Anthropic's analysis crew, described the corporate's measure of progress as "process horizon" — how lengthy an AI agent can work autonomously whereas enhancing the standard of its deliverables. "This time final 12 months, fashions might work for minutes," she stated. "Now, most of us have brokers working for hours on finish. Tomorrow, we'll have brokers which can be proactive, at all times on, and know what to work on with out dropping the body."
The occasion additionally included a number of infrastructure bulletins designed to assist builders preserve tempo. Anthropic stated it’s doubling its five-hour charge limits for Professional, Max, Crew, and Enterprise plans, and elevating API charge limits significantly. The corporate introduced a partnership with SpaceX to make use of the complete capability of its Colossus knowledge heart to increase compute availability — a direct response to the demand crunch Amodei described.
All three options are constructed into Claude Managed Brokers, which launched in public beta on April 8 as an opinionated harness that bundles finest practices together with reminiscence, device integration, and motion dealing with. Anthropic says groups utilizing Managed Brokers have shipped 10x sooner than these constructing their very own agent infrastructure from scratch. Albert described the platform utilizing an working system analogy: "With managed brokers, you don't want to consider all of the technicalities of the way you arrange the encircling system," he stated. "You're constructing an software for Macs — you don't wish to go should re-implement each element of macOS."
What dreaming, outcomes, and multi-agent orchestration imply for the way forward for enterprise AI
The aggressive implications are vital. As AI agent platforms from OpenAI, Google, and others compete for developer adoption, Anthropic is betting that manufacturing reliability — not simply uncooked mannequin intelligence — will decide which platform wins enterprise budgets. The dreaming characteristic specifically stakes out new territory: whereas different platforms supply reminiscence and gear use, the concept of brokers systematically reviewing their very own histories to extract reusable information goes additional towards the form of constantly enhancing methods that enterprises want earlier than delegating high-stakes work.
The convention showcased firms already working at that scale. Mercado Libre, Latin America's largest e-commerce platform, has 23,000 engineers working Claude Code, has reviewed greater than 500,000 pull requests with human oversight, and is aiming for 90% autonomous coding by the third quarter of this 12 months. Shopify has deployed Claude Code throughout not simply engineering however design, product, and knowledge science groups.
But it surely was Dario Amodei who articulated essentially the most expansive imaginative and prescient for the place all of this leads. He described a development from single brokers to a number of brokers to complete organizational intelligence — from "a crew of sensible folks in a room" to what he referred to as "a rustic of geniuses within the knowledge heart." And he reiterated a prediction he made roughly a 12 months in the past: that 2026 would see the primary billion-dollar firm run by a single particular person. "Hasn't fairly occurred but," he stated. "However we've bought seven extra months."
Dreaming is obtainable now in analysis preview. Outcomes and multi-agent orchestration are in public beta and accessible to all builders on the Claude platform. Whether or not seven months is sufficient time for a solo founder to construct a billion-dollar enterprise stays an open query — however after Tuesday, they’ve a couple of extra instruments to strive.
