Skip to content
Hire an AI employee
AI EmployeeSuper-AgentsAgent-to-AgentBenchmarksPricingBlog
Cluster 04 of 10

Multi-agent & AI organizations

Delegation, handoffs, org charts and orchestration — what changes when one AI employee becomes a team of them.

Multi-agent & AI organizations 50 articles
28 OCT 2026 Axiom Verifies the Finite Simple Groups Proof in Lean Axiom's AI agent swarms wrote a 43.1 million line Lean proof of the classification of finite simple groups in 24 days. They found 34 noteworthy deviations from the published record along the way. InsightsMulti-agent 6 min
Editorial data illustration on a near-white ground: a stack of paper labelled 15,000 pages, an arrow to a code window with a teal check and an orange 24 days tag, and the number 43.1M lines of Lean, under the headline Every finite simple group, checked
 24 OCT 2026 Avalon-M1: AI Agents Trained to Work as a Company General Reasoning's RLCR trains AI agents as one organisation with a shared reward. What emerged: lead engineers, peer review, more helping, and fewer reports to human overseers. InsightsMulti-agent 6 min
Editorial data illustration on a near-white ground: an org chart of small agent figures with a CEO at the top, a lead engineer in the middle and reviewers on the side, with a large number 9 and the words agents per team
 22 OCT 2026 Microsoft MAI-DxO in Science: 83% vs Doctors' 36% Microsoft AI's diagnostic orchestrator, published in Science, reached 83.2% on 304 hard NEJM cases against 36.1% for physicians. The five-role panel behind it, the costs and the limits. InsightsMulti-agent 6 min
Editorial data illustration on a near-white ground: five small teal figures sit around a navy case file at a round table, one figure holding an amber question card, with two big numbers, 83.2% for the AI panel and 36.1% for physicians
 22 OCT 2026 Anthropic Donates Agent Skills to the AAIF Anthropic donated Agent Skills, the SKILL.md standard used by 70+ agent platforms, to the Agentic AI Foundation: the name, GitHub org and domain move, and a security roadmap comes next. InsightsMulti-agent 5 min
Editorial data illustration on a near-white ground: a teal folder holding a SKILL.md card is handed from a navy Anthropic hand to an amber-roofed foundation building, while lines fan out from the folder to a grid of agent platform tiles labeled 70+ platforms
 21 OCT 2026 humans& Ampersand: A Model Trained to Know When to Speak humans& released Ampersand, a model built on GLM 5.3 and trained to sit in a group and choose between speaking and passing. What it scored, what it costs, and where it is weaker. InsightsMulti-agent 6 min
Data illustration on a near-white ground: a column of four chat bubbles, two tagged pass and one tagged speak with a teal line linking it back to the first, beside the large figures 17.3% for the base model with a partner and 39.0% for Ampersand with a partner
 21 OCT 2026 OpenAI Models Come to Google Cloud: What's Included OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are rolling out in private preview on Google Cloud, billed on the Google invoice. Microsoft stays the primary cloud partner. InsightsMulti-agent 4 min
Data illustration on a near-white ground: one teal model block sends three lines to three cloud shapes labelled Azure, Oracle and Google Cloud, the Google Cloud line in amber, beside the large text Oct 21
 21 OCT 2026 NVIDIA OpenShell: A Reference Architecture for Safe Agents NVIDIA's OpenShell Reference Architecture puts every agent in a kernel-level sandbox, runs it on Vera CPUs and polices it from BlueField DPUs. What shipped, and what is open. InsightsMulti-agent 5 min
Data illustration on a near-white ground: a small teal robot agent works at a laptop inside a glass box whose wall stops an amber arrow, standing on three layers labelled Rubin think, Vera act and BlueField govern, beside the large word OpenShell
 21 OCT 2026 KLS Conjecture Solved With AI: What the Proof Says Two researchers proved the 31-year-old KLS conjecture after exploring 100+ approaches with AI agents. Two more groups posted proofs within days, and unguided AI runs fell short. InsightsMulti-agent 5 min
Data illustration on a near-white ground: a branching tree of grey research paths with one teal path reaching a navy blob cut by a straight teal line, beside the large figure 31 and the lines years from conjecture to proof and 100+ approaches explored with AI agents
 21 OCT 2026 Daytona Raises $60M at $1B to Give Every Agent a Computer Daytona raised $60M at a $1B valuation to give every AI agent its own computer. We checked its growth figures against each other and priced a sandbox from its list. InsightsMulti-agent 7 min
Data illustration on a near-white ground: one small teal computer cube multiplies into a rising cascade of identical cubes, a few in amber, beside the large figures $1B valuation and 25x revenue growth since January
 21 OCT 2026 Claude Duet: Anthropic's Proactive Agent, Explained Anthropic opened a research preview of Claude Duet, an app where Claude briefs you, runs tasks on its own cloud computer and asks before it sends or pays. What it does and who gets it. InsightsMulti-agent 7 min
Data illustration on a near-white ground: a phone with one chat thread linked to a small cloud computer holding three task cards, one paused with an amber hold badge, beside the words Claude Duet, a proactive agent that checks with you first, 6,000+ connectors
 21 OCT 2026 ChatGPT Ads Hit $1B Run Rate: How the Ad Stack Works OpenAI says ChatGPT Ads reached a $1 billion run rate less than 200 days after launch. Its engineers explain the LLM that reads the chat, picks candidates and leaves the price to an auction. InsightsMulti-agent 5 min
Data illustration on a near-white ground: a column of chat bubbles feeding a teal orchestrator block that sends a few candidate cards toward an amber auction gavel, beside the large figure $1B and smaller text annual run rate
 21 OCT 2026 Anthropic API 2026-10-21: Breaking Changes and SDK 2.0 Anthropic shipped Claude API version 2026-10-21 and new SDK majors. Sampling parameters now return a 400, keys move to a Bearer header, and the SDKs switch to Apache 2.0. InsightsMulti-agent 5 min
Data illustration on a near-white ground: two API request cards, 2023-06-01 with temperature, top_p and top_k dials and a teal check, 2026-10-21 with the same dials crossed out in amber and a 400, beside a large 8 for seven SDKs and the CLI
 16 OCT 2026 GPT-6 Pro in ChatGPT Gets Parallel Agents: What's New OpenAI upgraded GPT-6 Pro in ChatGPT: faster responses, better plugin and tool use, and several agents working on parts of one task at once. What it does and what OpenAI left out. InsightsMulti-agent 4 min
Data illustration on a near-white ground: a navy chat window splits one task into three teal lanes, each with a progress bar, ending in three completed task cards, one amber with a check; the word PARALLEL above
 16 OCT 2026 Claude Designs Better Quantum Circuits: The Anthropic Paper Anthropic and Harvard researchers let Claude Mythos 5 redesign quantum circuits in a closed loop. It beat the best published designs on 7 of 8 tasks, one after a 53-hour run. InsightsMulti-agent 6 min
Data illustration on a near-white ground: a long dense quantum circuit with many teal gates, an amber arrow to a short clean circuit, and a huge 39x labeled lower qubit-gate cost, best of 8 tasks
 16 OCT 2026 Anthropic's Own Data: Agent Hours Outnumber Staff 6 to 1 Anthropic measured its own workforce: agents now log 6x the hours of the staff directing them, junior technical hiring fell, output rose, and human attention became the bottleneck. InsightsMulti-agent 8 min
Data illustration on a near-white ground: one person at a desk on the left, six teal agent terminals with amber clocks wired to the laptop, and a huge 6:1 labeled agent hours to staff hours, August 2026
 09 OCT 2026 Did Claude Discover ART? The Enzyme Dispute, Explained Anthropic's agents found ART, a CRISPR-like enzyme system, in 21 hours. A Copenhagen biologist says it matches his unpublished work. The receipts. InsightsMulti-agent 6 min
Editorial illustration on a near-white ground: a DNA strip with a highlighted run of repeats under a magnifying glass, a row of agent faces below, and the figures 950 Claude agents and 21 hours
 09 OCT 2026 Claude Dynamic Workflows: 1,000 Agents per Run Anthropic put dynamic workflows into public beta on October 9: one agent writes a program that runs up to 1,000 agents. What it does, its limits, what it costs and what is missing. InsightsMulti-agent 6 min
Editorial illustration on a near-white ground: one lead agent at a desk writing a plan that fans out into rows of small agent figures working in three phases, with the figures 1,000 agents per run, 64 at once and 66 of 70 bugs found
 08 OCT 2026 Claude Science Maps the Whole Sky in Ultraviolet Brice Ménard used Claude Science to merge 50 years of UV surveys and predict the missing third of the sky. Two agent reviews missed an artifact; he caught it. InsightsMulti-agent 5 min
Editorial illustration on a near-white ground: a sky globe split into measured tiles and dotted predicted tiles, with small agent figures filling the gaps and a magnifying glass over faint circles, with the figures 38,000 observations and about 10% error
 06 OCT 2026 Claude's 3SUM and APSP Algorithm: What the Paper Proves A paper by Alman and Vassilevska Williams credits Claude with the algorithm behind 3SUM in n^1.9992 and APSP in n^2.9995, refuting two core hypotheses. What it proves and what it does not. InsightsMulti-agent 8 min
Data illustration on off-white paper: a teal path runs through a thin crack in a tall charcoal wall, beside the large amber figure 1.9992 and the words new 3SUM exponent, down from 2
 01 OCT 2026 BootLoops: A Physicist's Open-Source Harness for AI Science Harvard physicist Matthew Schwartz released BootLoops, an MIT-licensed harness any model can drive, in an Anthropic guest post. What it does, what it found, and the caveats. InsightsMulti-agent 7 min
Editorial data illustration on a near-white ground titled BootLoops, with the numbers 36 manuscripts, 18 fields and 3 months, and eight colored islands labeled physics, ecology, genetics, linguistics, economics, cosmology, earth science and sunspots, linked to a central amber BootLoops block and wrapped in one navy outline
 29 SEP 2026 Intelligence Explosion Paper: What 22 AI Leaders Warn A working paper from 22 authors, including OpenAI's chief scientist and Anthropic's Jack Clark, asks what happens if AI automating AI R&D compresses years of progress into months. InsightsMulti-agent 6 min
Editorial data illustration on a near-white ground titled What if AI automates AI R&D: a teal staircase whose steps rise faster toward an amber top step, a teal feedback arrow looping from the top back to the bottom, and a large 10x labeled pace of AI progress in about 17 months in the paper's model
 25 SEP 2026 Claude Nine Loops: Anthropic's N=4 Physics Record One prompt, then keep going: Claude took a toy-model physics calculation from eight loops to nine on about $100 of compute. What Anthropic reported, who checked it, what stays open. InsightsMulti-agent 8 min
Flat data illustration on off-white paper: eight teal oval loops rising left to right like a staircase, a ninth loop on top in amber, a large number 9 labeled LOOPS, a server block labeled 96 CPUs 1 WEEK, a price tag labeled 100 dollars compute, and a note card reading KEEP GOING
 23 SEP 2026 Ten Claude Agents Proved a Faster Shortest Path in Lean Vals AI put ten Claude Opus 5.5 agents on a message board for 15 hours. They produced C-HD, a Lean-checked bound that beats Dijkstra in one narrow graph regime. What is proved, and what is not. InsightsMulti-agent 8 min
Editorial data illustration on a near-white ground: a corkboard with ten small agent badges pinned around a network map where one route is highlighted in rust, a large navy number 15 HOURS and a smaller 733 MESSAGES
 18 SEP 2026 Anthropic's Three Numbers: Claude Leads 26% of Its AI R&D Five days after the pacing essay, Anthropic published three numbers from inside its lab: Claude leads 26% of AI R&D, 1 in 47,000 agent actions blocked, 6% of R&D compute on safety. InsightsMulti-agent 13 min
Editorial infographic titled Three measurements from inside a frontier lab: a panel for AI-led AI R&D with 26% Claude leads and above 90% AI collaborates or above beside an AL0 to AL5 ladder marked at AL4; a funnel panel for oversight of agents running from 1 billion plus decisions through 100% pass the online monitor, 1 in 47,000 blocked, about 100,000 transcripts flagged a week and about 50 to human review a week, tagged about 30,000 agents at a time; and a compute panel with bars for 6% of AI R&D compute and 12% of AI-driven AI R&D compute on safety.
 17 SEP 2026 GitHub Rewrote Copilot in Rust: 800k Lines, One Developer GitHub published the receipts for rewriting its Copilot agent runtime from TypeScript into 832,378 lines of Rust: one developer, 128 pull requests, 14.5 weeks, about $120,000 in tokens. InsightsMulti-agent 11 min
Editorial infographic of GitHub's Copilot runtime port: a river of TypeScript pages on the left turning into a solid Rust block on the right, a counter reading 832,378 lines, a timeline from May to August 21 with 128 pull requests as ticks, a token meter at 136 billion, a price tag of about 120,000 dollars, and a stopwatch dropping from 5.25 seconds to 292 milliseconds
 17 SEP 2026 Claude Code Projects: One Chat Directs Parallel Threads Claude Code Projects now run from one conversation: a coordinator splits the work into parallel cloud threads with shared memory and keeps going when you leave. Beta for select Pro and Max users. InsightsMulti-agent 8 min
Editorial infographic of a Claude Code project: one conversation at the top labelled COORDINATOR, four parallel threads fanning out below it as cloud sessions on their own branches, a shared memory band running under all four, and a closed laptop at the side with the threads still running
 16 SEP 2026 Nous Research Cut 1M Lines of Code With 1,393 Agents On September 2, Teknium asked Hermes Agent to clean up a million-line repo. 1,393 subagents and 19 hours later the code was 34.4% smaller, for about $19,300. What the run showed and what it did not. InsightsMulti-agent 10 min
Editorial infographic of the Nous Research refactor: one orchestrator agent at the top of a three-level tree fanning out to 1,393 worker agents, with a counter showing non-test Python falling from 1,063,826 to 698,363 lines, a 19-hour clock, and a cost tag of about $19,300
 08 SEP 2026 Navier-Stokes: OpenAI's 10,000-Agent Proof and the Dispute OpenAI's Sept 8 post credits a group of roughly 10,000 concurrent agents, 2.7 million messages and 88 hours for a finite-time blow-up proof. Here is what the record shows, and what remains disputed. InsightsMulti-agent 17 min
Hand-drawn teal sketch of five dashed clusters of small dots connected by lines, labeled GROUPS and MESSAGES, one dot filled amber, beside a dashed box labeled SINGULARITY containing an inward spiral that stretches into a thin strand, and a clock labeled 88 HOURS
 31 AUG 2026 Meta's Project OT: The AI-Native Restructuring That Imploded, Explained Meta planned an AI-native company: pods of builders supervising agents, some teams shrunk up to 60 percent. Then incidents rose and the second wave was cancelled. The evidence ledger and the lesson. InsightsMulti-agent 7 min
Hand-drawn teal diagram of Meta's AI-native plan, a builder figure above a row of agent glyphs, a broken arrow circled in amber, and a rising incidents line labeled plus 40 percent with a clipboard labeled plan shelved
 30 AUG 2026 Claude Opus 5.1 Release Date: Leaks vs What Shipped Opus 5.1 and Sonnet 5.1 are community labels on two leaked test strings, not announced models. The dated fact-vs-rumor record of the marshmallow and melon leak. InsightsMulti-agent 9 min
Hand-drawn sketch of a solid box labeled Opus 5 next to two dashed boxes labeled Opus 5.1 and Sonnet 5.1 with amber question marks, and doodles of a marshmallow and a melon slice pointing at the dashed boxes with uncertain dotted arrows
 28 AUG 2026 Uber's AI Software Factory: 70% of Pull Requests, 3,600 Agent Skills, Flat Spend Uber published the most concrete enterprise-agent numbers yet: 70%+ of pull requests agent-attributed, 3,600 skills, 9.4x request growth on flat spend. The playbook, and the honest caveats. InsightsMulti-agent 7 min
Hand-drawn sketch of a factory labeled software factory with a conveyor of PR cards under robot arms labeled agents, a gauge at 70 percent, and a chart showing requests rising while spend stays flat
 19 AUG 2026 Fable 5.1 Is Out: Pricing, Benchmarks, and What Actually Changed Fable 5.1 shipped September 1, 2026. This tracker's rumor table is now a spec table: pricing held at $10/$50, cache reads fell 75%, and the long-horizon agent focus was real. InsightsMulti-agent 9 min
Hand-drawn timeline diagram from a solid box labeled Fable 5 to a second solid box labeled Fable 5.1 with a small amber checkmark above it and spec labels along the timeline
 31 JUL 2026 What Is a Super-Agent? Capability Breadth Without Category Hype A super-agent carries one goal across planning, tools, modalities, state, verification, and recovery in one loop. Breadth is a claim to test on a complete workflow, not a feature count. InsightsMulti-agent 26 min
Napkin-style sketch of one large agent loop circling icons for planning, tools, documents, data, code, and media, with a checklist gate at the loop's exit and an amber highlight on the verification checkmark
 31 JUL 2026 What Is Agent-to-Agent Communication? Discovery, Tasks, and Results Sending messages is not enough. A reliable exchange names the participants, task, authority, state, and result artifact - and knows when a plain API is the better interface. InsightsMulti-agent 29 min
Napkin-style sketch of two agent nodes exchanging a labeled task packet across a boundary line, with small stamps for identity, state, and artifact along the path, and an amber highlight on the task contract packet
 31 JUL 2026 Shared vs Role-Specific AI Memory AI agents should not share all memory. The safe default is isolation. Sharing is an explicit policy decision based on purpose, provenance, sensitivity, authority, freshness, and blast radius. InsightsMulti-agent 19 min
Napkin-style sketch of a central labeled bookshelf of approved sources with read arrows to three robots, each robot keeping its own small locked notebook, with an amber highlight on the single gated write arrow into the shelf
 31 JUL 2026 Multimodal AI Agents: When Work Crosses Text, Data, Code, and Media The value is not making text and images. It is continuity: a source becomes data, code validates it, a chart shows it, and the deck keeps the same facts. Every transition can lose meaning. InsightsMulti-agent 24 min
Napkin-style sketch of a chain of artifacts - a document, a data table, a code block, a chart, and a presentation slide - connected by arrows with small checkpoint gates between them, and an amber highlight on one checkpoint gate
 31 JUL 2026 Multi-Agent System Failure Modes: How Errors Propagate The first mistake may be small. The danger is propagation: another agent accepts the mistake as input, acts on it, stores it, or sends it farther. Assume one component will eventually be wrong. InsightsMulti-agent 28 min
Napkin-style sketch of a small crack in one box spreading along arrows through a chain of agent boxes, memory cylinder, and an external send icon, with an amber highlight on a circuit-breaker switch that cuts the chain
 31 JUL 2026 Manager-Agent Architecture: Routing, Review, State, and Failure Containment The manager is a coordination role, not an all-powerful agent. Identity, policy, budgets, state, and approvals stay deterministic - the controller proposes, services enforce. InsightsMulti-agent 25 min
Napkin-style sketch of a central manager node connected to three specialist worker nodes, surrounded by four small deterministic-service boxes labeled registry, policy, budget, and state, with an amber highlight on the approval gate between the manager and an external action
 31 JUL 2026 Manager Agent vs Specialist Agent: Different Jobs, Different Evals The manager succeeds when the right work reaches the right specialist and the outcome closes. The specialist succeeds when its artifact meets a domain contract. Do not score both with one number. InsightsMulti-agent 22 min
Napkin-style sketch of a manager robot at a routing desk with a task graph beside a specialist robot at a workbench with domain tools, each holding a different scorecard, with an amber highlight on the interface arrow between them
 31 JUL 2026 Human Span of Control for AI Agents: A Workload Model for Safe Supervision There is no universal number of AI agents one human can supervise. Agent count is an inventory number; human workload comes from the work each agent sends back. InsightsMulti-agent 24 min
Napkin-style sketch of a human figure at a desk with four labeled inbox trays for review, approvals, exceptions, and incidents, fed by arrows from several robot icons, with an amber highlight on a small reserve tank gauge beside the desk
 31 JUL 2026 How to Build an AI Organization Without Creating Coordination Debt The objective is not the largest agent org chart. It is the smallest human-AI operating model that produces more accepted work without moving the burden into delegation, review, and supervision. InsightsMulti-agent 22 min
Napkin-style sketch of a small org chart gaining one node while a dense fully connected web of agent boxes is crossed out, with an amber highlight on the single new connection
 31 JUL 2026 How AI Employees Work Together: Delegation, Handoffs, and Shared Context A useful collaboration has a beginning and an end: one role requests a defined result, another explicitly accepts it, the result arrives as a versioned artifact, and one owner closes the outcome. InsightsMulti-agent 22 min
Napkin-style sketch of two worker figures exchanging a stamped document across a labeled task pipeline with states for assigned, accepted, review, and closed, with an amber highlight on the acceptance stamp
 31 JUL 2026 Can AI Employees Manage Other AI Employees? Yes, when 'manage' means bounded coordination: decompose work, assign it, monitor state, review evidence, request repair, escalate exceptions. That does not make the AI manager an executive. InsightsMulti-agent 22 min
Napkin-style sketch of a manager robot at a small desk routing task cards to three worker robots while a human figure above holds a stop lever and an approval stamp, with an amber highlight on the escalation arrow to the human
 31 JUL 2026 Agent-to-Agent vs API Automation: When Should Software Delegate? Use an API when you can name the operation. Use an agent boundary when the outcome is clear but the path must flex. The transport does not decide - the contract scope does. InsightsMulti-agent 26 min
Napkin-style sketch of a fork in a path: the left branch leads to a plug-and-socket API icon with a typed parameter card, the right branch leads to an agent node holding a task envelope, with an amber highlight on the decision diamond at the fork
 31 JUL 2026 Agent Handoffs vs Agents as Tools: Who Keeps Control? The difference is not how many agents run. It is who controls the interaction and who owns the final result - and what happens to that ownership when something fails. InsightsMulti-agent 22 min
Napkin-style sketch of two panels: on the left a parent agent node holds a steering wheel while a small specialist returns a document to it; on the right the steering wheel is being passed across a dashed boundary to a second agent node, with an amber highlight on the steering wheel in transit
 31 JUL 2026 Agent Computer Use vs API Integrations: Reliability, Reach, and Risk A click is not proof of completion. Prefer the most structured action path that can complete the outcome, then verify the resulting state independently - through a different path. InsightsMulti-agent 22 min
Napkin-style sketch of an agent node with two action routes: an upper structured route through a plug-and-socket API icon straight to a database, and a lower winding route through a browser window with a cursor arrow, both converging on a verification checkpoint, with an amber highlight on the verification checkpoint
 31 JUL 2026 AI Organization Charts: Four Patterns for Human-AI Teams A box-and-line diagram that shows only titles and reporting relationships is incomplete. AI roles act across tools, share memory, and operate at machine speed - the chart needs operating overlays. InsightsMulti-agent 22 min
Napkin-style sketch of four small org chart patterns side by side - assistant per human, functional pod, manager with specialists, and shared service hub - with an amber highlight circling the human accountability marker on one chart
 31 JUL 2026 AI Agent Task Delegation: Authority, Acceptance, and Closure The delegating agent keeps responsibility for the parent outcome; the receiving agent owns only the contribution it explicitly accepts. A delivered artifact is not accepted work. InsightsMulti-agent 29 min
Napkin-style sketch of a parent robot holding a large parent task card while handing a smaller bounded child task card to a worker robot, the child card returning along a loop with an evidence stamp, with an amber highlight on the parent card staying in the delegator's hand
 31 JUL 2026 AI Agent Orchestration Patterns: Sequential, Parallel, Manager, and Peer Six patterns cover most deployed designs. The right one is the least complex topology that removes a measured bottleneck - and a single agent plus tools is always the baseline. InsightsMulti-agent 27 min
Napkin-style sketch of four small topology diagrams in a two-by-two grid - a straight chain of nodes, a fan-out of parallel branches rejoining, a hub with spokes to specialist nodes, and two peer nodes exchanging a task - with an amber highlight circling the hub diagram
 31 JUL 2026 AI Agent Handoff Protocols: What Must Travel With the Task An AI agent handoff should transfer a typed task contract, not a conversation summary. If the sender does not say whether ownership moves, both agents may act - or neither may own closure. InsightsMulti-agent 22 min
Napkin-style sketch of a robot handing a sealed packet across a gate to another robot, the packet labeled with small compartments for goal, state, sources, and authority, with an amber highlight on the acceptance gate latch
← The whole library