The Blueprint for Autonomous Work Agents | Gavriel Cohen, NanoClaw

Kovid Goyal, founder of NanoClaw, shared his experiences at AI Engineer Singapore, highlighting the unexpected viral growth of NanoClaw and its implications for the future of autonomous agents in enterprise.

The Serendipitous Encounter with Singapore’s Minister of Foreign Affairs

Goyal recounted a pivotal moment where Singapore’s Minister of Foreign Affairs (MFA) publicly detailed his extensive use of NanoClaw. The MFA had crafted a sophisticated “second brain” setup, integrating NanoClaw with Karpathy’s LLM Wiki, the Nemon memory system, and deploying it on a Raspberry Pi. This setup, which wasn’t part of NanoClaw’s original design, showcased the potential for advanced user customization and privacy-focused deployments. The MFA’s emphasis on privacy and security was a key draw, contrasting with larger, less transparent alternatives like OpenClaw.

Crystallizing Agent Adoption Strategies in Business

Goyal discussed two primary directions for agent adoption:

1. Agent Factories: Teams collectively building and managing agents to automate workflows.
2. Personal Agents: Individual users having their own assistant to augment their work.

The MFA’s personal “second brain” use case significantly clarified Goyal’s perspective: introducing agents to a company is best initiated by providing each person with their own agent. This approach acknowledges the learning curve involved in effectively interacting with agents (e.g., prompting, adjusting instructions, managing context windows). Goyal stressed that initial expectations of “throwing something in an agent and walking away” are unrealistic; successful agent utilization requires iterative fine-tuning of instructions and continuous interaction.

He identified the “second brain” use case—where users “dump in information” for the agent to collect, build internal memory (knowledge graph, LLM wiki), and provide useful outputs without expecting ready-made solutions—as the “killer use case” for autonomous agents today.

NanoClaw’s Security and Architectural Philosophy

Goyal contrasted NanoClaw with projects like OpenClaw, citing concerns over large codebases, numerous dependencies, and plain-text logging in OpenClaw that made it unsuitable for production environments or handling sensitive data. NanoClaw was built from the ground up with a focus on:

• Minimal Codebase: Starting fresh and being transparent for security review.
• Leveraging Existing Components: Utilizing robust tools like Agent SDK for core functionalities, reducing custom development and potential vulnerabilities.
• Isolation Model:
  • Running NanoClaw within a dedicated VM or Mac mini.
  • Isolating each agent in its own container runtime.
  • Crucially, no credentials are stored in the agent’s environment, preventing credential leakage even if prompt-injected.
  • All outbound requests from the agent are proxied through a vault, which conditionally adds necessary credentials.
• Access Policies and Human-in-the-Loop: Implementing granular access control, including mandatory human approval for sensitive actions (e.g., sending emails).

The Evolution to NanoClaw Co. and Enterprise Deployments

NanoClaw’s unexpected viral success, particularly after a tweet from Karpathy, led Goyal to pivot from an AI-native marketing agency to focusing entirely on NanoClaw. The new company, NanoClaw Co., aims to address the demand from executives and CEOs who have successfully built personal agent setups and now want to deploy them across their organizations.

NanoClaw Co. offers enterprise deployment services, often on the client’s own infrastructure (cloud or on-premise). This involves complex integrations with existing enterprise systems, including:

• Credential management (SSO, vaults)
• Observability platforms
• Security systems
• Internal data sources

Goyal emphasized that while many companies have strong DevOps and security teams, they often lack the specialized “AI engineering” expertise needed to confidently deploy and manage agents. NanoClaw Co. positions itself as a partner, providing this expertise for initial setup and ongoing management.

A key distinction of agents from traditional enterprise software is their constant evolution. Unlike software that can run for years without updates, agents require continuous upgrades (e.g., to new LLM versions, new features) to remain effective and competitive. NanoClaw Co. will assist in this ongoing maintenance.

Philosophical Questions and the Future of Open Source with AI

The discussion also delved into broader questions about the future of development workflows:

• Git/GitHub for Agent Factories: Goyal pondered if traditional Git/GitHub models are ideal for agent-driven development, especially for internal discussions within open-source projects. He suggested that agent interactions might be better suited for platforms like Slack, with GitHub serving as a final merge point.
• Messaging Platform Tier List: Kovid shared his preference for Slack, followed by Discord/Teams, then Telegram, with WhatsApp and personal phone number integrations being least preferred due to limitations.
• Managing Open Source Projects with AI: A significant challenge identified is the exponential increase in pull requests generated by coding agents, making triage and review extremely difficult. Goyal proposed moving towards “prompt requests” instead of code, and a “wiki-driven development” model. In this model, all future developments, feature requests, and bug reports would reside in a central wiki. Agents would pull context from the wiki for tasks and push back updates, effectively buffering and structuring the development process.

• Tags:
• AI Agents
• Autonomous Agents
• Feature
• Enterprise AI
• AI Engineering
• Agent Security
• Second Brain
• NanoClaw
• LLM Wiki
• Open Source Management
• Nemon
• Karpathy