Agent Runtime

Build, run, and audit AI agent workflows on Tanqory.

The Agent Runtime is in design. This page documents the shape we are building toward, not a shipped product.

What it is

Most agent demos run a prompt, call a tool, and end. Production commerce needs more: retries, audit, memory across turns, human approval gates on destructive actions, and per-tenant cost budgets.

The Tanqory Agent Runtime is the execution environment for those workflows.

Design

  • Orchestrator + specialists — one planner agent decomposes a goal, domain-scoped agents execute. Specialists in v1: Commerce, Marketing, Support.
  • Tool calling contract — agents only call MCP tools (see /mcp). Same auth, same audit, same scopes as a human developer.
  • Memory — hot (Redis) for conversation, warm (Postgres) for episodic, cold (S3) for archived runs, vector (MongoDB Atlas) for semantic recall.
  • Human-in-the-loop — destructive operations pause for approval. The runtime resumes from the exact step after the operator clicks accept.
  • Observability — OpenTelemetry traces every step. LangFuse (or similar) surfaces prompt-level cost and latency, not just app-level.

What we will not do at v1

  • We will not ship eight "specialist agents" out of the gate. We will ship three (Commerce, Marketing, Support) and iterate.
  • We will not promise autonomous "Build a business" from one prompt. Humans approve at every destructive step until the model proves reliability.
  • We will not build a bespoke orchestrator. We will adopt a production framework and earn the right to replace it later.

Roadmap

  • Design-partner preview: Q3 2026
  • Commerce Agent v1: Q3 2026
  • Marketing + Support agents: Q4 2026
  • Public GA: Y2