// PRODUCT_02

VAFI

Viloforge Agentic Fleet Infrastructure

Deploy AI coding agents like infrastructure. Executor pods clone repos, implement code, run tests, and commit. Judge pods verify independently. Scale your engineering output like you scale containers — without scaling headcount.

Core Capabilities

  • check_box Autonomous executor and judge agents running as Kubernetes pods
  • check_box Same container image for both roles — configured via environment variable
  • check_box Shared workdir so judges review exactly what executors built
  • check_box Full execution traces captured in CXDB for every task
  • check_box Three-layer Docker image: base → claude → agent
  • check_box Harness agnostic — swap AI providers without changing specs or workflows

Specifications

architecture:
GitLab Runner-like
controller:
Python asyncio
harness:
pluggable (Claude Code, Codex, custom)
runtime:
Kubernetes pods + shared volumes
images:
harbor.viloforge.com/vafi/*
traces:
CXDB per-task capture
deploy:
Helm chart

Harness Agnostic

VAFI separates orchestration from AI. The controller manages the lifecycle — claim, clone, execute, test, report. The harness is a pluggable interface. Today it runs Claude Code. Tomorrow it could run Codex, Aider, or your own custom harness.

Your task specs, review gates, and execution traces stay the same regardless of which AI does the work. Every agent session is captured as a replayable trace — debug any task by replaying the exact conversation. No vendor lock-in. Ever.

terminal Get Early Access

helm upgrade --install vafi charts/vafi/ --namespace vafi-prod

vafi-controller-executor-01
00:00:01[BOOT] vafi-controller v0.8.3 starting...
00:00:02[BOOT] Role: executor | Agent: executor-01
00:00:03[BOOT] VTF endpoint: https://vtf.viloforge.com
00:00:04[BOOT] Session volume: /sessions/
 
00:00:07[POLL] Checking vtf board for claimable tasks...
00:00:08[CLAIM] TASK-047 "Implement JWT refresh token rotation"
 
00:00:11[SETUP] Creating workdir /sessions/task-047/
00:00:12[SETUP] Cloning vilosource/auth-service @ main
00:00:14[SETUP] Writing .vafi/context.md (spec + history)
 
00:00:17[HARNESS] Invoking Claude Code...
00:00:18[HARNESS] Prompt: "Read .vafi/context.md and implement the task"
00:00:20[HARNESS] Claude Code working...
00:01:45[HARNESS] Session complete — 3 files modified, 47 tool calls
00:01:46[HARNESS] Trace captured → CXDB task-047-exec-001
 
00:01:49[TEST] Running gate: pytest src/tests/
00:02:05[TEST] 38 passed, 0 failed
 
00:02:07[REPORT] Submitting completion to vtf...
00:02:08[REPORT] TASK-047 → pending_completion_review
 
00:02:11[POLL] Checking vtf board for claimable tasks...
00:02:12[WAIT] No tasks available. Retry in 30s... _