# Barnacle Labs — Full Reference > London-based AI consultancy helping enterprises build production AI agents, sovereign AI systems, and AI strategy. We engineer breakthroughs, not buzzwords. Founded in 2022 by Duncan Anderson (former European CTO, IBM Watson AI) and John-David Wuarin (MS Engineering, Columbia). We build AI systems for the US Federal Government, financial services, global enterprises, and biotech firms. --- ## About Barnacle Labs Duncan and JD got early access to GPT-3 before ChatGPT launched — before most people had heard of generative AI. Duncan spent years as European CTO for IBM Watson, leading AI implementations across enterprise. JD built open source projects in Duncan's team that scaled to production systems. When generative AI went mainstream, they knew how to build with it, not just talk about it. The name comes from Darwin, who spent 8 years meticulously studying 10,000 barnacles. That detailed, unglamorous work provided crucial evidence for the theory of evolution. Barnacle Labs builds the meticulous systems — infrastructure, architecture, production code — that enable scientific and commercial breakthroughs. ### What We Do - **AI Discovery** — Workshops + working prototype in 2 weeks, delivered as an AI Action Plan - **Build** — Prototype-first development through to production shipping - **Training** — AI training for leadership, developers, and business teams - **Advisory** — Architecture, models, and deployment guidance at critical decision points ### Our Bets 1. **Agentic AI** — Autonomous agents with memory that act on your behalf, reason through multi-step problems, and collaborate with each other across enterprise and scientific workflows. 2. **Sovereign AI** — Small, specialised models that work in constrained, secure environments where data can't leave your walls. 3. **Domain-specific tools** — Built for the realities of particular problems. --- ## AI Agents Guide URL: https://www.barnacle.ai/ai-agents-guide A comprehensive strategic guide for enterprise leaders, distilling Barnacle Labs' production experience building AI agents. ### Key Concepts **What are AI agents?** AI that reasons, uses tools, and pursues goals adaptively — not just chatbots. Agents design their own steps rather than following pre-defined workflows. **The Agentic Loop:** Observe → Think → Act → Repeat. The agent perceives its environment, reasons about what to do, takes action using tools, and evaluates the result. **Workflows vs Agents:** Workflows follow pre-defined steps orchestrated by code. Agents decide their own steps based on the situation. Workflows are predictable and auditable; agents handle ambiguity and novel situations. **Identifying Agent Opportunities (5 criteria):** 1. Skilled people spending time on low-value work 2. Tasks that happen frequently 3. Work that requires judgement but follows patterns 4. Data is digitally accessible 5. Stakes are manageable if the agent makes mistakes **Case Studies:** Lead generation, legal research, brief writing, customer support, research monitoring, drug discovery, contract analysis, customer due diligence, ontology generation, newsletter generation. **Key Risks:** Security (prompt injection, data leakage), hallucination (confident but wrong outputs), cost (token usage at scale), bias (training data reflected in outputs). Each with specific mitigations. **Getting Started:** Pick a real problem, choose technology (framework recommendations at /resources/agent-frameworks), build tools the agent can use, build evaluation systems before scaling. --- ## Agent Memory Guide URL: https://www.barnacle.ai/agent-memory-guide A guide to the concepts that matter for AI agent memory, explaining why vector databases alone aren't enough. ### Key Concepts **The Problem:** Most AI agents start every task from zero. The context window is volatile — everything disappears when the session ends. Institutional knowledge is locked in formats agents can't access or reason over. **Why Vector Databases Aren't Enough:** Vector similarity ≠ relevance. Chunks lose surrounding context. Vectors encode no relationships, time, or structure. The agent doesn't control what gets retrieved. Every memory is equal in the eyes of cosine similarity. **Layers of Memory:** - Working memory — the context window (volatile, limited) - Long-term memory — episodic (what happened) + semantic (what's true) - External memory — documents, databases, APIs **The Retrieval Problem:** Score memories by Recency, Importance, and Relevance (Stanford Generative Agents model). Then navigate — graph traversal (BFS with max steps) or agent-based retrieval where a lightweight agent reasons about the question and navigates the memory graph. The top-level agent can request more if needed. **Hybrid Retrieval (Table Stakes):** Run keyword search and semantic search in parallel, then rerank results. Cuts retrieval failures roughly in half. This is baseline — necessary but not sufficient for reasoning-based retrieval. **Graph Memory:** Store entities and relationships, not just chunks. Enables multi-hop reasoning. Community detection algorithms help graph memory scale. **Context Graphs:** Store the full reasoning context around decisions — not just what was decided, but why, what alternatives were considered, and what evidence supported it. **Memory Creation:** A background process that watches everything — conversations, tool outputs, external events — and creates raw memories continuously. Captures first, judges later. The consolidation process handles organisation and understanding. **Consolidation (Learning From Sleep):** Two passes — shallow consolidation extracts basic facts and relationships in near real-time; deep consolidation runs periodically to cluster, abstract, resolve contradictions, and build understanding. Neither pass should destroy raw material. **Never Delete:** Supersede old memories rather than deleting them. The history of a decision — how it evolved, why it changed — is more valuable than just the latest state. **Memory in the Weights:** LoRAs and memory layers can embed frequently-used knowledge directly into model weights — swappable cartridges that eliminate retrieval latency. The boundary between what the model knows and what it retrieves is blurring. **Barnacle Labs is building a production memory system as part of the Minerva scientific reasoning platform, with particular interest from financial services organisations. They are seeking design partners.** --- ## Sovereign AI Guide URL: https://www.barnacle.ai/sovereign-ai-guide A practical guide to AI privacy, compliance, and data sovereignty for UK/EU organisations using US AI providers. ### Key Concepts **The "Stateless" Assumption:** Many organisations assume that because LLM APIs are stateless (no session persistence), there's no data protection issue. This assumption fails for four reasons: 1. **"Stateless" isn't** — Providers retain request logs for up to 30 days. The NY Times vs. OpenAI case demonstrated that prompts can be discoverable in litigation. 2. **US jurisdiction** — The US CLOUD Act (2018) allows US authorities to compel data from US companies regardless of where the servers are physically located. "It's in London" isn't a privacy strategy. 3. **UK GDPR** — Sending personal data to a US model provider is likely a "restricted transfer." The UK-US Data Bridge only covers organisations that actively opted in. Compliance depends on Biden-era Executive Order 14086. 4. **Political risk** — EO 14086 and PCLOB oversight face uncertainty under changing US administrations. **Three-Tier Response:** 1. **Tighten hygiene** — Treat prompts as personal data, minimise what goes to models, understand retention terms 2. **Add anonymisation** — Replace sensitive data with placeholder tags cached locally before sending to models. Barnacle Labs offers a production-ready anonymisation layer proven in healthcare. 3. **Sovereign compute** — Move to EU/UK-controlled infrastructure. European options include Mistral AI, Scaleway, OVHcloud, Nscale. --- ## Your Employees Are Already Running AI Agents URL: https://www.barnacle.ai/openclaw-guide A guide to OpenClaw, the fastest-growing open-source project in history, and the ecosystem of clones, forks and competitors reshaping enterprise AI. ### Key Concepts **What is OpenClaw:** A free, open-source AI agent built on four primitives — persistent identity (SOUL.md), periodic autonomy (heartbeat), accumulated memory, and social context. Runs locally, connects to any LLM via API key, integrates with messaging platforms. **The Moltbook Experiment:** 32,000 agents joined a social network in 48 hours. They created forums, shared discoveries, and founded a religion (Church of Molt) that spread through SOUL.md rewrites. Demonstrates emergence at scale and the risks of writable identity files. **Security Reckoning:** 21,000+ exposed instances, CVE-2026-25253 (CVSS 8.8), 824 malicious skills on ClawHub (8% of ecosystem), 22% of enterprises have shadow deployments. Architectural issue: broad permissions + natural language control + prompt injection = enormous blast radius. **The Clone Ecosystem:** NanoClaw (container isolation), ZeroClaw (Rust, 3.4MB, best memory), Moltis (enterprise, sandboxed), PicoClaw (edge/IoT), NullClaw (678KB Zig binary), memU (memory layer), HiClaw (multi-agent), NemoClaw (NVIDIA enterprise governance). **Enterprise Playbook:** Assess shadow deployment exposure. Separate agent runtime from inference layer. Start with contained read-only use cases. Build skills governance early. Watch memory and orchestration layers. **Barnacle Labs builds production agent systems for enterprises. If your team is evaluating agent deployment or discovering shadow usage, talk to us.** --- ## Minerva Research Platform URL: https://www.barnacle.ai/minerva Barnacle Labs' scientific reasoning platform. Unlike fully autonomous AI scientist approaches, Minerva keeps human researcher intuition in the loop throughout — researchers invoke and interact with agents step-by-step, not just review outputs at the end. **Capabilities:** - Navigate the research landscape and identify promising directions - Test hypotheses by simulating protein behaviour with AlphaFold - Spin up compute to train models on genomic and proteomic data - Run parallel AI conversations that branch and merge while maintaining context - Collaborate with teams, sharing workflows and datasets **Status:** Early prototypes running. Targeting biotech, pharma, academic, and federal health organisations. --- ## NanCI URL: https://www.barnacle.ai/nanci AI-powered cancer research app built by Barnacle Labs in partnership with the US National Cancer Institute (NCI) and Google Cloud. Uses Google Gemini AI. - Deployed at 100+ academic institutions worldwide - Available on the Apple App Store - Recommends research papers personalised to each researcher's interests - Learns from user behaviour to proactively suggest relevant papers - Chat with papers for instant answers - Topic folders and bookmarking - Researcher social networking and event discovery - Addresses the problem of 500+ cancer research papers published daily Dr. Oliver Bogler, Director of the Center for Cancer Training at NCI: "We wanted to build a solution that would automate literature curation with an integrated social networking component to bring joy back into the scientific process." --- ## Blog Posts URL: https://www.barnacle.ai/blog | Title | Author | URL | |---|---|---| | Building a Production Multi-Agent System for Document Writing | Barnacle Labs | /blog/2026-01-16-multi-agent-document-writing | | Seatbelts for AI: Lessons from the Grok Image Controversy | Barnacle Labs | /blog/2026-01-19-seatbelts-for-ai | | Building AI products means managing API costs. Here's what actually worked. | Barnacle Labs | /blog/2026-02-02-building-ai-products-means-managing-api-costs-here | | OpenClaw and the Programmable Soul | Barnacle Labs | /blog/2026-02-02-openclaw-and-the-programmable-soul | | The Longest Chain | Barnacle Labs | /blog/2026-02-07-the-longest-chain | | The Missing Piece of the Claude Code Workflow: Isolated Worktree Databases | Barnacle Labs | /blog/2026-02-07-the-missing-piece-of-the-claude-code-workflow-isol | | "It's in London" isn't a privacy strategy for US AI | Barnacle Labs | /blog/2026-02-25-your-ai-vendor-s-london-region-won-t-save-you | | Why We're Going All-In on Forward Deployed Engineers for AI Projects | Barnacle Labs | /blog/2025-07-15-fde | | Stop Treating AI Like Magic: Why Context Engineering Beats Bigger Models | Barnacle Labs | /blog/2025-07-15-what-is-context-engineering | | Barnacle Labs Appoints Dr. Oliver Bogler as Scientific Advisor | Barnacle Labs | /blog/2025-07-23-oliver | | The 4 Levels of AI Agents: When to Use Workflows vs Autonomous Systems | Barnacle Labs | /blog/2025-09-25-agents-intro | | It's not in the stdlib — Prompt like a data engineer | Barnacle Labs | /blog/2025-09-27-pater-on-ce | | Understanding Cancer and Telomerase: From Biology to New Treatments | Kevish Napal | /blog/2025-11-01-kevish-1 | | Building a Reproducible Multimodal Pipeline | Kevish Napal | /blog/2025-11-07-kevish-2 | | Understanding and Processing CT Imaging for Stroke Detection | William Auroux | /blog/2025-11-12-william-1 | | The Dimension Dilemma: Why 2.5D Models Outperform 3D CNNs for Stroke Classification | William Auroux | /blog/2025-11-19-william-2 | | Mean Pooling Beats Attention: Predicting Telomerase Activity from Whole-Slide Images | Kevish Napal | /blog/2025-12-18-kevish-3 | | Beyond Data Hoovering: The Nuanced Reality of Training Large Language Models | Barnacle Labs | /blog/2023-06-19-beyond-data-hoovering | | Me and an AI: Building a New Website Together | Barnacle Labs | /blog/2025-01-20-website | --- ## Agent Framework Recommendations URL: https://www.barnacle.ai/resources/agent-frameworks Production-tested recommendations, updated quarterly: - **Claude Agent SDK** — For building reasoning agents - **n8n** — For workflow agents in a low-code environment - **Vercel AI SDK** — For simple agents in TypeScript --- ## Contact - Website: https://www.barnacle.ai - Careers: careers@barnacle.ai - LinkedIn: https://www.linkedin.com/company/barnacle-labs ## Sitemap https://www.barnacle.ai/sitemap.xml