A dual agentic loop
Keep talking while your agent works. The conversation loop handles exchanges, retrieves useful context and follows ongoing work. It can assign a task to an agent or start a process such as n8n. The execution loop handles background work: reasoning, using tools, checking results and delegating subtasks.
The conversation controller remains internal; only tasks use the harness assigned to the agent. Before execution, Galaris resolves the model, effort, context, permissions, tools and policy. The harness receives a bounded envelope without secrets; its stream contains events followed by exactly one terminal result. Voice has its own state machine for latency and interruptions.
Understand the two loops · Controllers and harnesses
Agents that collaborate
Delegation creates a persisted subtask, linked to its origin. The requesting agent can follow its progress and use its result to continue the mission. Research, calculation and review can be assigned to agents with different skills and tools.
Collaboration has explicit boundaries: team membership does not grant full management rights over other agents’ tasks. Each agent retains its permissions, model profile and task harness. Organise an agent team.
Documents coauthored by humans and agents
Deliverables become addressable, revisable and shareable documents. Humans and agents work on the same canonical document, with its history, sources and permissions. Its reference can travel between conversations and delegated tasks; corrections enrich the shared work.
HTML documents and JSON Datasets form a knowledge base and can support applications. The browser SDK galaris.datasets connects these applications to data with access checks and versioned consent. A task’s Working Set distinguishes sources, versions, artifacts, destinations and receipts: claiming file delivery requires verifiable delivery evidence.
Documents and applications · APIs and Dataset SDK
Hybrid memory, maintained by Dream
Continuity combines recent history, lexical and semantic recall, provenance and durable memory. ACLs apply before similarity ranking. Galaris builds task context from authorised memories and a versioned Working Set; the conversation controller builds its own context. Resources travel through canonical URIs.
When enabled, Dream maintains memory during available periods: selective extraction, matching, consolidation and topic organisation. Experience reflection uses task evidence to retain a procedure, correction or pitfall. This learning enriches the knowledge agents can reuse; it does not retrain model weights.
Memory, documents, files and business objects retain separate responsibilities: a memory does not replace its source or the permissions protecting it. Memory and recall · Dream mechanisms.
Decision models such as Jev
Galaris combines generative LLMs with a Decision resource, including TypeSafe’s Jev. The model receives a state, questions and permitted options, then returns a structured choice. This specialisation complements models that write, code and summarise. OpenRouter’s introduction to Jev.
| Mechanism | Bounded decision | What follows |
|---|---|---|
| Dispatcher | Select a route / effort pair allowed by the harness and policy. | Execute directly or use a plan. A single permitted choice avoids a model call. |
| Topic classification | Reuse a candidate thematic dossier or request a new topic. | Organise activities into knowledge spanning conversations. |
| Memory retention | Ignore, link to existing memory or request extraction. | Add provenance or ask the text model to extract facts worth keeping. |
The contract validates returned questions, selected options and probability distributions when supplied. The profile controls whether a bounded, recorded text fallback is allowed for recoverable unavailability; this fallback never invents confidence scores. The effective model and fallback reason remain inspectable. Galaris retains control over actions and permissions.
Models, profiles and harnesses
Work that survives interruptions
Tasks, attempts, execution leases, waits and results are persisted in PostgreSQL. The scheduler can resume work after interruption; a question for a human becomes a durable wait. Long-term goals advance through task cycles, with a separate judge examining evidence and progress.
| Recovery mechanism | Guarantee and scope |
|---|---|
| Leases and attempts | Identify who is executing the task and track attempts through to its result. |
| Internal harness effect checkpoints | Journal calls with durable effects before and after execution. An unknown outcome blocks automatic replay. |
| External processes | Retain a run identifier, idempotency key, events and link to the task. |
| External harnesses | Share the task contract; qualify their recovery and isolation capabilities separately. |
Tasks and orchestration · Goals · Processes
Observable and evaluable AI
Identifiers connect tasks, attempts, model calls, tools, processes, messages and deliverables. Inspect decisions, errors and available costs, distinguishing known, estimated and incomplete costs. Traces provide the context needed to understand an execution.
The AI Lab turns tasks, text conversations and voice conversations into versioned datasets. It compares Dispatcher, Planner, Task/Conversation/Voice executors and Dream/Goal mechanisms on frozen cases, with semantic rubrics and resumable runs. A weak output, a candidate failure and a judge failure remain distinct. The interface and tools restricted to authorised agents let you operate these evaluations.
Explore the Lab · Supervision and diagnostics
An open architecture under your control
The stack combines FastAPI, Pydantic, SQLAlchemy, PostgreSQL and pgvector, a Vue and Quasar interface, and an internal Pydantic AI harness. Docker supports deployment. The six layers of the agentic OS separate infrastructure, providers, inference, agents, shared services and applications.
Profile APIs and the LLM gateway serve your clients; Janus provides agent access from a compatible client. Durable inferences add frozen requests, attempts and replayable streams. MCP exposes authorised tools: the catalogue intersects user, agent, connection, function and execution-context permissions, including during deferred discovery. Messaging channels, external harnesses and processes such as n8n connect to this foundation.
Explore the compatible ecosystem
APIs and integrations · Inventory of 244 native tools · Compare approaches
An accessible setup
A simple Docker installation, a preconfigured application, then your OpenRouter API key: you can start talking to the Galaris assistant. The initial profile selects resources by usage, including Jev for decisions. You can then adapt models, providers and tools to your needs through the graphical interface.
Operations include health checks, logs and harness lifecycle actions. Availability and capabilities depend on the selected configuration; qualify backup, restoration and load on your deployment.
Deployment and data control · Configure models and harnesses

