Agent infrastructure for Fabric, an agentic coding IDE

Context

Current work at Farpoint on Fabric (codewithfabric.com): a sovereign, agentic AI coding IDE that plans first and codes second. My focus is the agent infrastructure, the layer that lets the model invoke tools, hold context across steps, and act reliably on large codebases.

Problem and constraint

Text-only agents edit code the way a person pastes into a single file: no real model of the project, no durable context, no coordination across files. That breaks down on large, multi-file work and on SQL that has to stay correct. The job was to give the agent structured, observable access to the codebase and tools without every call becoming an unbounded side effect.

Approach and key decisions

• Architected Model Context Protocol (MCP) infrastructure for structured tool invocation and multi-step agent orchestration, so capabilities are explicit, typed tool calls rather than free-form text.
• Built tool-abstraction and execution layers for secure LLM-to-system interactions, with the boundary observable: a call that cannot proceed becomes a structured signal the agent can reason about.
• Implemented context-lifecycle and memory pipelines for long-running autonomous workflows, so the working context stays relevant over many steps.
• Designed distributed processing pipelines for large-scale codebase analysis and transformation, plus iterative reasoning loops for autonomous code improvement beyond traditional code completion.

Outcome (sanitized)

Agents operate on real repositories through structured tool calls, persistent context, and distributed analysis instead of one-file text patches. Concrete metrics are under NDA; happy to walk through specifics in an interview.

What I'd do differently

Invest in the observability stream even earlier. The decisions that mattered most only became obvious once each step of the loop was inspectable.