Development roadmap

This project is being built in three deliberate stages, in this order:

1. Separate — stand the engine up as its own repo/module that compiles on its own (no dependency on the private product). In progress.
2. Make it usable in Lexelligence — the private product consumes it as a Go module dependency, with all Lexelligence-specific content kept private (see open-core overlay).
3. Generalize & publish — decouple the last product-specific seams, ship a neutral example domain + great docs, tag v0.1.0, and open it to contributors.

We treat the engine’s DB-as-program model (a declarative workflow stored in Postgres), the two-tier semantic cache, and the rules-first / LLM-second execution model as the differentiators. Other Go agent frameworks (e.g. AgenticGoKit) are builder/code-first and don’t emphasize caching or a deterministic rules layer — we borrow good ideas from them (below) without copying that shape.

Stage A — Separate (v0.1.0 target)

• Run scripts/extract.sh to carry the engine packages + rewrite imports.
• Decouple from the product: replace loopcontroller’s websocket broadcast and vector storage with small interfaces so the engine is transport- and framework-agnostic.
• go build ./..., go vet ./..., go test ./... all green locally.
• Ship docker-compose.yml (Postgres + pgvector) + a Makefile (migrate, test, example).
• Neutral, runnable example domain (replace any regulatory seed).

Stage B — Usable in Lexelligence

• Lexelligence go.mod: require + local replace => ./agentic-golang.
• Swap Lexelligence’s in-tree internal/agentic/* for the module import; delete the duplicated copies.
• Lexelligence keeps only its domain pieces (EU-AI-Act workflow seed, expert prompt content, API/auth/transport glue).
• scripts/publish.sh proven for the “push to both repos” flow.

Stage C — Generalize & publish (v0.2+)

Feature milestones (several inspired by AgenticGoKit and Kunal Kushwaha’s “Building Agentic AI Systems in Go” — credited, not copied):

Milestone	Feature	Notes / inspiration
v0.2	Structured-output validation gate	Validate model JSON vs. the agent schema; reflective repair loop bounded by max_iterations
v0.2	prompt_templates loader + render	Per-agent expert prompts (persona + few-shot); mechanism generic, content pluggable
v0.3	Tool calling + MCP	Let agents call Go functions and discover tools via Model Context Protocol — AgenticGoKit
v0.3	Orchestration patterns	First-class sequential / parallel / DAG / loop / sub-workflow over the existing graph — AgenticGoKit
v0.3	Tiered observability	minimal / standard / detailed levels; optional OTLP exporter alongside the Postgres sink — AgenticGoKit / Kushwaha
v0.4	Mermaid workflow export	Render a session’s executed graph as a Mermaid diagram — AgenticGoKit
v0.4	Eval framework	Semantic matching + LLM-as-judge with confidence scoring, building on the existing hallucination proxy — AgenticGoKit
v0.4	Streaming-first	Stream tokens through the loop to subscribers — AgenticGoKit
later	Builder API	Optional code-first agent/workflow construction layered over the DB definitions, for users who don’t want to seed SQL — AgenticGoKit

Open-core overlay (using it in Lexelligence)

The engine reads its “program” (workflows, agent definitions, prompt templates, model/retry/cache policies) from the database and from an injected config struct — it has no compiled-in domain knowledge. Lexelligence therefore supplies its specifics without modifying the engine:

• Lexelligence’s own private migrations/seeds provide the EU-AI-Act workflow and the expert prompt content.
• A thin wiring layer in Lexelligence constructs the engine services (modelrouter, semanticcache, observability, …) with product config and a concrete Broadcaster.

For local engine development that needs Lexelligence-shaped data, put it under agentic-golang/local/ — that path is git-ignored in this repo (see .gitignore), so product-specific material never gets pushed to the public repo even though you edit and push the engine from the same directory.