System I · Production-Proven
A hybrid AI agent program built from two components: the Fable-5 multi-agent framework (forked and customised) and the DeepSeek-v4-pro API. Four specialist sub-agents, persistent SQLite memory, and a three-phase pipeline that has shipped dashboards, websites, maps, and reports.
Architecture
Veles is not a single agent — it is an autonomous software engineering team. The Orchestrator plans and delegates. The Architect reads and proposes. The Coder writes and runs. The Verifier audits and certifies. All powered by DeepSeek-v4-pro.
Sub-Agent Roles
Each sub-agent has a distinct access level, purpose, and speed profile. The separation of concerns — planning, analysis, implementation, and verification — is what makes the hybrid approach auditable and reliable.
The lead agent. Decomposes the task, assigns subtasks to specialist sub-agents via compact design packets (goal, constraints, owned paths, acceptance criteria, verification commands), and synthesises their outputs. Never writes code directly. Makes the judgment calls.
Explores the codebase, reads relevant files, and proposes concrete, code-ready improvements. Cannot write or execute — only observe and recommend. Perfect for design audits, enhancement planning, and codebase exploration before committing to changes.
The implementer. Writes files, runs commands, executes the architect's plan or the orchestrator's design packet. Has full read/write access within the workspace. Implements changes, runs tests, and reports results — but does not self-certify.
The auditor. Runs tests, checks evidence, and returns PASS/FAIL with line references. Has no stake in the implementation — its only loyalty is to correctness. Every change is verified before it ships. Catches what the Coder missed.
“The Coder implements. The Verifier audits. They are never the same agent. This separation of concerns is not a nice-to-have — it is the structural guarantee that makes autonomous software engineering defensible.”
Veles Architecture Principle
Pipeline
Every Veles project follows the same three-phase pattern. The phases are sequential but iterative — the Verifier can send any phase back for revision. Nothing ships without a PASS.
veles new <project> creates the directory, seeds configuration, and writes initial memory. Reasonix then builds the project using its full creative context and skill pipelines — strategy, design, implementation. Veles provides the scaffold; Reasonix provides the craft.
veles design <path> "query" invokes the Architect for codebase exploration and improvement proposals. Reasonix implements the changes. This is the polish phase: expand, refine, iterate. The Architect catches what the builder missed.
veles verify <path> runs the Verifier — a full quality audit returning PASS/FAIL with evidence and line references. All FAIL items are fixed. When every check is green: git push. Deployed. Done.
The pipeline is linear by design — Bootstrap → Enhance → Ship — but the Verifier can send any phase back for revision. A FAIL in Phase III means returning to Phase II. A design flaw caught in Phase II means revisiting Phase I. This loop is not a bug; it is how quality compounds.
Memory
Veles uses SQLite for durable project memory — facts, decisions, data sources, and design constraints persist across sessions. When the Orchestrator delegates work, sub-agents receive a compact design packet seeded from memory. No context bloat. No re-discovery. The system remembers what it learned last time so you don't pay to re-learn it on every run.
Configuration
Veles auto-detects DEEPSEEK_API_KEY from the Reasonix environment. All configuration lives in fable.toml. No manual wiring required.
# fable.toml — Veles configuration [fable] provider = "openai-compatible" model = "deepseek-v4-pro" base_url = "https://api.deepseek.com/v1" memory_backend = "sqlite" max_iterations = 20 max_delegations = 8
| Command | Purpose | Speed |
|---|---|---|
veles status | Health check — all systems green? | < 1s |
veles new <project> | Bootstrap a new project with config and memory | ~5s |
veles design <path> [query] | Architect reviews code, proposes improvements | 5–15s |
veles verify <path> | Verifier audits quality, returns PASS/FAIL | 10–30s |
veles run <path> <task> | Full autonomous pipeline: delegate → implement → verify | 30–90s |
veles memory <path> | View project memory | < 1s |
| Project | Type | Outcome |
|---|---|---|
| Electric Cooking Market Intelligence | Dashboard | Interactive market dashboard with Chart.js, dark slate theme, amber accent, KPI cards, technology comparison table |
| Clean Cooking Transition | Informational Site | 7-section narrative scrolling site with hero, stat counters, chart visualizations, technology cards, and timeline elements |
| Clean Cooking Report | PDF Report | 5-page data-driven report with title page, executive summary, data tables, policy recommendations, source attribution |
| Cybersecurity Landscape Intelligence | 4-Tab SPA | 4-tab single-page application with KPI grid, Chart.js visualizations, ransomware trends, NIST CSF framework analysis |
| World Map of Economies | Interactive Map | Folium-based interactive world map with choropleth layers, regional drill-downs, KPI overlays, and data filtering |