Skip to main content

Architecture and build plan

This is the plan the platform was built from

It records the design as decided before any code was written, including code sketches that were never compiled. Where the built platform differs, Project status says how, and the rest of this site describes what was actually built.

Open-Source Back-Office Agent Platform: Architecture and Build Plan

Stack: Rust, Loco.rs, CopilotKit, YAML-first configuration, OSI-approved licenses only.

This is the single consolidated plan. It merges the original architecture plan with the Loco.rs and YAML-first revision. Facts marked (source-verified) were read on 10 Nov 2026 from the published loco-rs 1.3.1 and graph-flow 0.8.1 crate sources or from repository LICENSE files. Code marked // VERIFY uses an API that was not confirmed against source, and // TODO(impl) marks code to complete during implementation. Nothing in this document has been compiled.

TL;DR​

  • What it is: a platform of back-office agents for small and mid-sized businesses. The first workflow is invoice processing: ingest a document, extract fields, validate, match against records, get human approval, post to accounting, audit every step.
  • Backend: one Loco.rs application (Apache-2.0) providing the HTTP server, config, database, Postgres job queue, scheduler, auth and CLI. Framework-free library crates sit behind traits.
  • Configuration: everything that describes behaviour lives in YAML. Workflows are YAML files compiled at boot into graph-flow graphs from a registry of Rust building blocks called task kinds.
  • Frontend: CopilotKit v2 (MIT) talking AG-UI 1.0 over SSE to the Loco app. The agent list and the approval card are driven by the same YAML.
  • Models and documents: vLLM serving Qwen3.8-27B and Qwen3-Embedding-0.6B (both Apache-2.0). Docling-serve (MIT code) parses documents.
  • Defaults chosen for license reasons: native Rust connectors and OpenTelemetry Collector + Jaeger. Valkey replaces Redis and SeaweedFS replaces MinIO in optional services. The core needs neither.
  • Effort: about 23–32 engineer-weeks across eight phases, starting with an automated license gate in CI.

Sections​