Atlas

The platform layer for AI-native software.

Atlas makes AI systems operable through explicit boundaries, typed contracts, declared dependencies, private transport, gateway-routed agents and models, and evidence that can be reviewed.

Abstract Atlas platform architecture with connected agents, applications, data, and security controls

Capabilities

Contracted service boundaries

Capabilities own one domain and expose it through OpenAPI contracts. Current examples include conversations, blobs, transcription, connections, search-mail, send-mail, Moltbook, radio charts, and personal library.

Surfaces

Focused applications

Surfaces declare the capabilities, agents, routes, and environments they need. Runtime binding happens through scoped context, not hidden imports.

Agents

First-class actors

Agents run through a gateway with persistent threads, structured outputs, tool access, declared permissions, and a path toward traceable work records.

Models

Registered inference

Model assets are declared separately from applications and served through a model gateway, so a surface can consume inference without owning model runtime details.

Workflows

Durable steps

Workflows coordinate capability and agent calls through a ledger. Each step can record inputs, outputs, judgments, quality signals, and failure state.

Evals

Behavioral evidence

Unit tests cover deterministic code. Evals cover agent and workflow behavior where the question is quality, relevance, or judgment.

Operating model

Declared state, reconciled systems.

  • Manifests: apps, workers, workflows, capabilities, agents, models, routes, placement, and dependencies are declared.
  • Contracts: OpenAPI and models drive generated clients, runtime dispatch, and policy checks.
  • Environments: local, dev, and prod are separated by state, secrets, routes, and deployment targets.
  • Private networking: Tailscale makes host placement flexible while preserving explicit service boundaries.
  • Release evidence: checks, tests, evals, artifacts, and deployment records are becoming part of the platform record.

Why it matters

AI needs platform discipline.

AI-native systems need shared answers for tools, memory, identity, prompts, evaluations, routing, deployment, observability, and provenance.

Atlas turns those concerns into platform primitives that can be tested, deployed, reviewed, and improved independently.

Current engineering focus

Traceability, source-backed memory, and workload identity.

The next platform layer is evidence. Atlas is moving toward trace IDs across surface requests, capability calls, workflow steps, model inference, agent turns, tool calls, evals, and deployments. Memory is being treated as source-backed platform state rather than raw transcript recall. Workload identity is the security layer that lets surfaces, workers, workflows, agents, and gateways act with bounded authority.