Skip to main content

​
Narrative

Most agentic coding tools burn LLM tokens at every layer: planning, routing, execution, validation. The FORGE + TOMMY architecture inverts this pattern. FORGE is the only layer that uses intelligence — it reads GitHub issues, understands context, and generates structured task payloads. TOMMY is pure mechanical execution: it receives pre-structured tasks and runs them via SSH + Claude Code CLI on a VPS. When FORGE delegates to TOMMY, the orchestration layer itself consumes zero additional LLM tokens. The intelligence is front-loaded into FORGE’s single planning pass.

​
Key Messages

  • “Intelligence once, execution many” — FORGE thinks, TOMMY acts. No redundant LLM calls in the execution loop.
  • “Your GitHub issues are your backlog AND your execution queue” — FORGE reads epics/blocks/tasks directly from GitHub. No separate task management layer.
  • “Auto-close on green” — When TOMMY reports ALL_PASS, FORGE closes the GitHub issue with a summary comment. The developer returns to a clean backlog.

​
Metrics to Track

  • Token cost per resolved issue (FORGE planning pass only vs. full agent-loop alternatives).
  • Time from trigger to issue closure (end-to-end autonomous cycle time).
  • Success rate (ALL_PASS vs. HAS_FAILURES ratio across different repo types).

​
Target Audience

Developer tooling teams evaluating autonomous coding agents. The pitch: you don’t need an expensive agent loop for execution — you need good planning and mechanical reliability.

​
Status

This is a potential campaign narrative, not yet validated with external audiences. The architecture works but has known limitations (Vercel timeout ceiling, credential re-binding friction). Recommended next step: internal demo with metrics before external positioning.