LangChain Paris meetup: agents in production, with Pigment
LangChain × Pigment
- What I kept
- Tools are turning into code generated for each request. Pigment dropped its purpose-built analysis tools: the agent now writes Python that runs in a sandbox and calls MCP itself. Tools stop being a fixed catalogue and become code produced for the request.
- What it changed
- Starting point of the second phase of pythagoras-sec and of its on-demand connectors: instead of a catalogue of prebuilt MCP tools, the connector is generated as code for the request, so the engine has to secure code written milliseconds earlier rather than a fixed list of tools.
Evening meetup on what it takes to make agents reliable in production. Harrison Chase (co-founder and CEO, LangChain) opened on the state of agent development, the architectures LangChain sees in the field and why teams should own their stack. Thomas Delayen (Pigment) followed with how Pigment builds agents, using Analyst, their finance analysis agent, as the case study.
What Pigment showed
- One MCP server, several clients. Claude, ChatGPT and Pigment’s own agent reach the same MCP server. Isolation is layered: tenant isolation at the gateway, permissions per tool, row-level access rights on the data, an audit trail tenants can read and monitoring kept internal.
- From tools to code. Early 2026, the agent called three purpose-built tools (contribution, variance and time-series analysis). Mid 2026, it writes Python scripts that run in a sandbox and call MCP themselves; the dedicated tools are gone, and a virtual filesystem holds the larger scripts.
- Their rules for building an agent now: deep agent pattern, code over tools, compare yourself to other harnesses, invest in skills and tools and use libraries for the glue, breadth of features first.