The future of AI isn't just about better chatbots — it's about creating virtual collaborators that can work alongside humans as capable, autonomous partners. At HubSpot for Startups' annual AiSummit in San Francisco, Matt Bell from Anthropic delivered a visionary keynote exploring how AI systems are evolving from simple assistants into powerful collaborators that multiply human capabilities.
Matt Bell, who works at Anthropic, shared insights from one of the most pioneering companies in AI, outlining the roadmap for virtual collaborators and how startups can integrate these capabilities into their products.
The Asymmetric Advantage of Human-AI Collaboration
The key to effective virtual collaboration lies in understanding the complementary strengths of humans and AI systems. Claude demonstrates remarkable advantages in specific areas:
- Speed and comprehension: Claude reads 500 times faster than humans with superior comprehension
- Writing velocity: Significantly faster content generation capabilities
- Knowledge access: Encyclopedic world knowledge at instant recall
- Task execution: Tireless execution of both simple and complex tasks with assistance
Meanwhile, humans excel in areas where AI still struggles:
- Long-term focus: Maintaining coherent goals and priorities over extended periods
- Strategic planning: Complex planning and execution capabilities
- Common sense reasoning: Better intuitive understanding of real-world contexts
- Self-correction: More balanced correction without under or over-correcting
- Domain expertise: Maximum quality output in specialized areas of knowledge
The Evolution from Assistant to Pioneer
Bell outlined a clear trajectory for AI development over the coming years:
2023-2024: The AI Assistant Era - Chat-based interactions where users ask questions and AI responds in conversational format.
2025: The AI Collaborator Era - AI acts as a force multiplier, enabling users to produce more work, better work, and faster work. Users become significantly more capable employees through AI assistance.
2026-2027: The AI Pioneer Era - Teams of AI agents working together will conduct frontier work, including scientific breakthroughs and building complex systems under human direction.
This evolution is supported by reliable scaling laws. Research from METER shows that AI systems have exponentially increased their ability to handle complex tasks — from handling tasks that take humans seconds to complete, to minutes, and now approaching hour-long tasks. This trend is expected to continue over the next two years.
Core Building Blocks of Virtual Collaborators
The Agentic Loop
Virtual collaborators operate through an autonomous planning and execution cycle. Unlike rigid workflows, agents excel at open-ended problems without clear predetermined processes. The agent receives a user request, accesses available tools, and runs a continuous feedback loop of thinking, tool usage, result analysis, and task completion assessment.
Bell demonstrated this with an insurance claims example, where an agent processes a reimbursement request by determining goals, following procedures, pulling records, analyzing against policy, and proposing actions — all with minimal human intervention.
Access to Data and Tools
Models are only as effective as the context and tools available to them. Anthropic developed the Model Context Protocol (MCP), an open standard that enables AI applications to connect seamlessly with third-party tools and data sources.
Before MCP, connecting agents to complex APIs like Google Workspace required extensive custom development work. MCP standardizes these connections, offering benefits for:
- Developers: Easy integration with multiple third-party services
- Tool vendors: Accelerated adoption through standardized AI compatibility
- End users: Better applications with broader functionality
- Enterprises: Clearer boundaries and standardized interconnects across AI initiatives
The protocol has gained significant momentum, with major companies including OpenAI and Google adopting the open standard.
Intelligent Indexing and Search
Effective virtual collaborators require organized, accessible knowledge systems. Traditional Retrieval Augmented Generation (RAG) approaches often fragment information into disconnected snippets that lack context.
Anthropic developed Contextual RAG, which has Claude write descriptions that contextualize knowledge snippets, making them understandable as standalone pieces. Combined with other innovations, this approach reduced failed retrievals by approximately 3x.
Agentic search proves far more powerful than single-shot retrieval. Claude performs iterative searches — using early searches to understand problems and later searches to dive deep. This enables answering complex, multi-part questions across large knowledge bases.
Anthropic uses this internally across Slack, Google Docs, and other sources, with Claude building knowledge graphs to orient itself to company activities. Bell personally uses this system 10-20 times daily for company orientation, team identification, and document feedback.
Memory Systems
Virtual collaborators need memory to handle long tasks that exceed context windows and to learn from repeated task execution. Anthropic implements two primary memory forms:
Context Compression: When approaching context limits, Claude introspects on its context window, identifying the most important information to preserve. It compresses nearly 200,000 tokens by 5-10x while maintaining essential information for ongoing work.
Memory Files: Claude takes notes on tasks for future reference, often creating complex directory structures. Testing with video games like Pokémon demonstrates this capability — Claude creates organized files covering game mechanics, locations, and NPC interactions, even developing meta-protocols like "getting unstuck procedures."
Coding and Virtual Machines
Large language models and software capabilities complement each other perfectly. Tasks where models struggle — complex mathematics, handling massive datasets — are ideal for code-based solutions.
Anthropic optimizes each Claude version for software engineering using SuiteBench as their North Star benchmark. This benchmark requires making changes to complex codebases to implement features that pass hidden tests.
Virtual machines unlock entirely new workflows:
- Code development: Building, modifying, and testing codebases safely
- Single-use software: Enabling non-programmers to prototype interfaces through natural language descriptions
- Data science: Processing large, complex datasets through code generation
- Document editing: Successive edits rather than constrained start-to-finish writing
Multi-Agent Systems
Multiple Claude instances working together significantly outperform single instances. Bell drew parallels to human civilization—while modern humans have the same neural capacity as Stone Age ancestors, better organization, technology, and cultural transmission enable extraordinary collective achievements.
Anthropic's Advanced Research exemplifies multi-agent capabilities, tackling complex search questions requiring hundreds or thousands of searches. An orchestrator delegates subtasks to specialized sub-agents, including dedicated citation specialists, enabling comprehensive research projects beyond single-agent capabilities.
Multi-agent systems excel in:
- Search and retrieval: Comprehensive research across multiple domains
- Coding projects: Learning codebases before implementing changes
- Map-reduce tasks: Breaking large problems into manageable components
- Tool specialization: Managing hundreds of tools through specialized sub-agents with focused tool sets
Safety and Responsible Implementation
Virtual collaborators' expanded capabilities require robust safety frameworks. Key considerations include:
Permission Management: Virtual collaborators should operate within the same permissions as their human users, not with blanket super-user access.
Human Oversight: Avoiding "approve button mashing" through interfaces that require genuine understanding of AI actions.
Responsibility Models: Clear ownership structures for virtual collaborator actions and outcomes.
Attack Vectors: Protection against data poisoning from untrusted sources and potential data exfiltration risks.
Mitigation strategies include careful access management, read-only tool defaults, sandboxed environments, human sign-off requirements, automated guardrails using additional Claude instances for monitoring, and controlled internet access.
Getting Started with Virtual Collaborators
For startups looking to integrate virtual collaborator capabilities, Anthropic provides comprehensive resources at docs.anthropic.com, including tutorials, courses, reference materials, and cookbooks. The platform features an AI assistant that can search available knowledge to answer specific implementation questions.
Organizations can immediately begin leveraging virtual collaborator capabilities through Claude.ai and Claude Code integration, tools that Anthropic uses extensively internally for team acceleration.
The virtual collaborator represents a fundamental shift in how we work with AI — from asking questions to delegating complex, multi-step projects. As scaling laws continue their reliable trajectory, virtual collaborators are positioned to become a meaningful fraction of global GDP within two to five years, potentially making every knowledge worker twice as productive and creating trillions in economic value.
~~~~~
AI Disclaimer: AI helped us summarize the key points from these videos, and our editorial team reviewed everything to make sure it's clear and correct.
