A multi-agent platform I built and run myself, across a four-node fleet, mostly on local models. It handles commerce, sales, trading research, and content work, and I talk to it by voice or chat.
Where there is a number below, there is a script that reproduces it.
- 4-node compute fleet, specialized by role (control/business, two GPU nodes, support).
- ~55 agents and 85+ services, discovered through a single registry.
- 54 local language models served via a hot-swapping gateway, plus cloud models.
- MQTT event bus (60+ event namespaces) wiring every component together.
- Cumulative vector memory: 640,000+ points across 12 collections.
- Real-time voice stack, full observability, and self-healing.
These figures are measured on the live fleet: see metrics/snapshot-2026-06-20.md for the exact commands and their outputs.
The control surface: a real-time view of every node, agent, pipeline, and KPI, with a built-in voice and chat assistant.
A few of the operational panels:
Commerce & marketing. Async integrations with Shopify, WooCommerce, Stripe, Printful, dropshipping suppliers, and TikTok Shop, plus ML engines for pricing, demand forecasting, recommendation, customer-lifetime-value, fraud detection, sentiment, product enrichment, and visual search. Storefronts, catalogs, and campaigns run with little to no human input.
Autonomous sales. A pipeline with prospection and qualification live, and negotiation and closing built, backed by a deal ledger and a compliance guard. Self-hosted e-signature and full end-to-end validation are in progress.
Trading. A research-and-execution stack running in paper/testnet mode: strategy models, time-series forecasting, and a reinforcement-learning world model that simulates business and market decisions before they are taken.
Creative & content. Long-form and social content generation, SEO articles, community management, image and video generation, and scheduled multi-channel publication.
Voice & multimodal assistant. A real-time conversational assistant drives the platform by voice or chat: streaming speech in and out, speaker identification, vision, and document understanding (OCR, parsing). Ask it for a status, a report, or an action and it routes the request to the right agent.
Builders. Generators that produce web apps, games, and storefront themes, plus self-improving dev pipelines (audit / build / architect) that scan, generate, and refactor code and infrastructure under execution verifiers.
A layer above the agents tracks how the system is doing and adjusts it. An analytics hub, coupled to the Command Center, consolidates KPIs, finance, and ML predictions and can act on agents directly. Two prediction stages run live: a Chronos-class time-series forecaster and a DreamerV3-class world model that simulates the outcome of a decision before it is taken, trained on the platform's own history. A causal-reasoning stage is in progress. Underneath that, a self-evaluation loop: validators, QA with self-healing tests, a learning engine using A/B canaries and evolutionary search, and drift monitoring. A skill manager turns recurring gaps into new reusable capabilities. Long-horizon planning and a research engine are being added.
Four complementary systems, not one store: a structured knowledge wiki (raw to wiki to index, link graph, full-text search, note lifecycle) compiled by a librarian; an episodic memory of intents, decisions, errors, and reflections with graceful decay; a shared, versioned strategic-state graph used by the cognitive modules as common ground; and a retention manager that ages, archives, and purges under time-to-live policies. A vector knowledge base and embeddings give every agent durable, searchable context.
Command center & dashboard. A control plane classifies each request, routes it to the right agent and model, manages lifecycle, and surfaces everything in a live dashboard with the built-in voice/chat assistant. A mobile controller and Notion sync extend control beyond the desk.
Research & knowledge. An autonomous research agent gathers and synthesizes external sources; a knowledge backbone (embeddings, a librarian, vector search, document ingestion) gives every agent durable, searchable context.
Reliability & observability. An autonomous reliability layer (SOC plus SRE) detects anomalies, correlates them, drives incidents through a state machine, and remediates from runbooks with LLM-assisted investigation, on top of a metrics/logs/alerting stack and data-governance services (schema registry, validators, retention, aggregation), plus defensive security and compliance monitoring. Business-support agents handle finance, accounting, legal, and customer service.
Talos is organized into planes: a control plane (orchestration, routing, lifecycle, dashboard), a knowledge plane (a four-part memory of knowledge wiki, episodic store, strategic state, and retention, plus research and skill generation), an execution plane (business, creative, and trading agents), a cognitive layer (a stack from prediction and governance to shared state, with long-horizon planning and meta-research being added), and a reliability plane (observability and self-healing). An MQTT event bus and a local-model gateway connect them across the fleet.
Python (async-first) · local LLM ops (llama.cpp / GGUF, model routing, VRAM-aware hot-swap) · multi-agent orchestration · graph-RAG memory · vector database · MQTT event bus · real-time voice (streaming STT/TTS, speaker ID) · Prometheus-compatible metrics, logs, and alerting · reverse-proxy + WAF ingress · systemd · CI with ruff, pytest, and pre-commit.
- Execution-verified generation over single-shot prompting.
- Constrained decoding for valid output; ensemble judges to catch semantic false positives.
- Defense-in-depth: multiple independent checks, never one prompt.
- Metrics-driven: every change is measured; if it isn't reproducible, it isn't done.
Parts of the infrastructure are extracted, generalized, and released as standalone libraries:
- agent-resilience - circuit breaker, Redis-backed DLQ, and offline MQTT buffer.
- async-api-client - resilient async REST client with rate limiting, retries, and pagination.
- multi-agent-orchestrator - capability-based task routing template.
The platform itself is private; the components above are the parts released under MIT.