Agents in the Sandboxes — LangChain Community London × Focused.io
Agents in the Sandboxes
What happens when an AI agent goes rogue?
LangChain Community London × Focused.io
AI agents are becoming increasingly capable.
They can reason, use tools, call APIs, browse the web, execute code, interact with enterprise systems, and take actions on our behalf.
But as we give agents more autonomy, another question becomes just as important:
What happens when an agent does something it was never supposed to do?
Join LangChain Community London × Focused.io for Agents in the Sandbox, an evening exploring how we design, control, observe, and safely operate agentic AI systems in the real world.
We’ll look beyond the hype and focus on the engineering challenges that appear when agents move from demos into production.
An agent might call the wrong tool, follow a malicious instruction hidden inside external content, access something outside its intended scope, enter a costly loop, or take an action that should have required human approval.
The challenge is not simply making agents more intelligent.
It is making them powerful enough to be useful, while controlled enough to be trusted.
What We’ll Explore
Throughout the evening, we’ll discuss some of the most important topics shaping production agentic AI:
Agent sandboxing and controlled execution
Tool permissions and least-privilege access
Prompt injection and untrusted content
Trust boundaries across agents, tools, APIs, and data
Human-in-the-loop approval for high-impact actions
Agent observability and tracing
Evals, guardrails, and behavioural testing
Failure modes, retries, loops, and unexpected actions
Moving from agent prototypes to reliable production systems
We’ll also explore where technologies such as LangChain and LangGraph fit into the architecture of modern agentic applications.
At the centre of the event is one key question:
How do we give AI agents enough autonomy to be useful without giving them more freedom than we can safely control?
Schedule
6:30 PM — Doors Open
Food, drinks, networking, and conversations with the community.
7:00 PM — Welcome: Agents in the Sandbox
Why agent control, security, and observability are becoming critical as AI systems gain more autonomy.
7:15 PM — [TBD Talk 1]What Happens When an Agent Goes Rogue?
Exploring real-world agent failure scenarios, including unexpected tool use, prompt injection, excessive permissions, and unintended actions.
7:45 PM — [TBD Talk 2] Building the Sandbox
A technical look at permissions, trust boundaries, guardrails, observability, and human oversight.
8:15 PM — [TBD Talk 3] From Demo to Production
What changes when an agent moves from an experiment into a real production environment.
8:40 PM — Community Q&A and Discussion
Bring your questions, architectures, experiences, and opinions.
9:00 PM — Closing & Networking
Who Should Attend
This event is for:
AI and ML engineers
Software engineers
Solution and platform architects
Security practitioners
Product managers and AI product builders
Founders and startup teams
Researchers and students
Developers working with LangChain or LangGraph
Anyone interested in agentic AI and production AI systems
You don’t need to be a LangChain expert.
If you’re interested in AI agents, production engineering, AI security, observability, or responsible autonomy, this event is for you.
About LangChain Community London
LangChain Community London brings together developers, AI engineers, researchers, founders, and technology enthusiasts building applications with large language models and AI agents.
The community explores technologies and engineering practices across LangChain, LangGraph, agentic systems, RAG, evaluation, observability, and production AI, with a focus on practical learning, collaboration, and sharing real-world experience.
About Focused.io
Focused.io is a software and AI engineering company that works alongside product and engineering teams to build production-ready systems.
Focused.io works across areas including custom AI agents, production RAG systems, custom software development, and eval-driven development, with a strong emphasis on reliability, testing, and moving AI systems from prototype to production.