Zebric Agent

Technical preview

@zebric/agent is not yet a stable v1 release. The public surface and deterministic harness are implemented, but distribution, compatibility policy, and some operational capabilities remain in progress.

Zebric Agent is a generic agent for Zebric applications. It discovers the skills published by a running application, turns supported OpenAPI operations into typed tools, and can deterministically validate Blueprints inside a configured workspace.

It does not contain issue-tracker, QA, or browser-automation business logic. The issue board is the reference scenario: its Blueprint publishes the QA vocabulary and workflows, while an external executor performs the actual tests.

Agent API and Zebric Agent

The two components are complementary:

Component	Responsibility
Agent API	Application-owned contract, authentication, scopes, state transitions, jobs, idempotency, and audit events
Zebric Agent	Model orchestration, contract discovery, typed tool creation, local approval policy, retries, and Blueprint validation

Zebric Agent builds on LangChain Deep Agents internally, but Deep Agents types are not the stable Zebric-facing contract.

Supported preview use cases

• Validate a TOML or JSON Blueprint without a model.
• Connect to one or more running Zebric applications.
• Discover and call bounded read operations.
• Opt into exact semantic mutations through a callback or human-in-the-loop approval.
• Observe asynchronous workflow jobs and resume an interrupted mutation safely.
• Run deterministic contract, real-runtime, and compiled-CLI tests without a real LLM.

Read Getting started, then Approvals and mutations. Application authors should begin with Build an Agent-Drivable Application.