The engine behind Hextant.
A multi-agent orchestration engine. It dispatches parallel agent sessions, runs interactive multi-turn conversations with the Hextant executives, and chains workflows across Claude Code, Anthropic, OpenAI, and Ollama under one governance substrate. The infrastructure under every Hexaxia AI tool that talks to more than one model.
How do dozens of parallel agents stay coordinated?
Background dispatch
Fire tasks at agents, collect structured results.
Drop a prompt against a target, or fan out a template across N targets in parallel. The dispatcher tracks budget, retries on transient failure, and returns one aggregated payload. Designed for the work that does not need a chat: status checks, code reviews, audit passes, scaffold runs.
Interactive sessions
Streaming multi-turn with mid-conversation delegation.
Open a session against an executive (vCIO, vCFO, vCMO, etc.), get streaming responses as SSE, and let the executive-hub delegate to specialists mid-conversation. The session remembers context. Cost is tracked per session, queryable any time.
Workflows
Chained multi-step, dependency-ordered.
Compose dispatches into a workflow. Independent steps run concurrently. Each step’s output feeds the next via named placeholders. Audit a tool with one command. Build a new agent end to end (types, implement, tests, review) with another. Save the full result tree to disk.
The system in six layers.
Each layer is replaceable. The PAL router treats providers as adapters; new ones drop in without touching dispatch code. The hub treats sessions as a protocol, not a runtime. The event bus keeps the pieces loosely coupled on purpose.
CLI surface
`python -m orchxia` for dispatch, hub launch, workflow runs, template lookup. The way operators actually drive the engine day to day.
Hub server
FastAPI on port 8300. REST endpoints for dispatch and templates, WebSocket for streaming, SSE for session messages. The integration surface for UIs and downstream services.
Provider adapters
Claude Code CLI, Anthropic API, OpenAI, Ollama. Unified interface so dispatch code stays provider-agnostic and routing decisions stay declarative.
Core engine
17-state dispatcher, PAL (Provider-Adapter-Layer) router with three tiers, stagnation detection. The reliability layer that keeps long-running parallel work from silently stalling.
Event bus + aggregator
Pub/sub event bus for delegation, broadcast, and progress signals. Aggregator with five strategies for fanning out one request to many agents and collapsing results back to one payload.
RhizomeRAG proxy
Native bridge to RhizomeRAG so sessions can ground responses against the shared knowledge graph without each agent reimplementing retrieval.
Productized tasks.
Templates turn a recurring agent task into a single command. One run does the same operation across every executive, with structured output you can pipe into a workflow or save to disk. The set ships against the Hextant tools today; new templates land as the tool catalog grows.
The same substrate as Hextant.
Every dispatched agent runs against AGF, APF, and ASIP. AGF defines what the agent is allowed to do on this dispatch. APF defines how it behaves while it does it, including stagnation detection and self-fix. ASIP defines how the agent itself evolves, with every change routed through a human review queue before it lands.
That is why orchestrating dozens of parallel agents does not produce dozens of independent drift trajectories. Authorization, behavior, and change control are not per-agent decisions. They are substrate. See the frameworks.
Built for teams running their own AI stack.
Engineering leaders who are past the prototype stage. Teams that need to dispatch and aggregate work across many models without writing a new orchestrator every quarter. Operators shipping production AI workflows who want the reliability layer (state machine, router, stagnation detection) without building it themselves.
Not built for: teams that need one-off agent chat. Not built for: closed vendor stacks where the routing decisions are already made for you. Not built for: places that treat AI as a chatbot, not as a workhorse.
See the engine before you commit to the stack.
OrchXia is in active development. Clients on Fractional CAIO and Build-with engagements get access to the current version and the roadmap. Tell us what you are trying to orchestrate and we will say whether OrchXia is the right answer.