Agentic Framework
Build agents for workflows where failure isn't an option. Give them tool access, safety boundaries, and a full audit trail, not just a prompt.
Agents retrieve from trusted sources. Pre-built registry connectors for clinical trials, drug interactions, medical coding, and more. MCP for tool servers you bring yourself.
Typed inputs, explicit tool schemas, guardrails around action-taking. Agents can pause for human approval and resume only after consent.
Replayable traces and OpenInference export. Every decision and tool call recorded for debugging, compliance, and quality assurance.
Multi-turn conversations grouped by context. Strict isolation between patient encounters, calls, or workflows. No shared state.
They can't reach the data they need at runtime, so they guess. They can't act safely on external systems, so they stall. The Agentic Framework gives agents typed connectors for data and a controlled execution layer for actions, with traces for everything.
How it works
Receives requests, plans steps, selects and invokes connectors, returns results as tasks or messages. Contexts carry state across turns.
Five types. Registry for pre-built capabilities, MCP for external tool servers, Agent for nested agents, A2A for remote agents, Schema for custom tools.
Conversation state with strict isolation between contexts. OpenInference trace export for debugging and observability.
User-facing assistants
Agents that answer questions, edit documents, and drive admin workflows, connected to your knowledge base instead of guessing from training data.
Summarize the key risks in this incident report and draft a mitigation plan.
incident-report-2026.pdfAutomated pipelines
Trigger agents by API or event to process records, generate reports, and update downstream systems. Monitor every step with tasks and traces.
Customer-embedded agents
Let customers bring their own tools and systems. Agents combine Corti capabilities with third-party and customer-owned services through typed connectors.
Pull the latest test results for patient pat_8421 and check against current guidelines.
Compliance and audit workflows
Agents that can check policies, enforce rules, and log every action. Full traceability from request to output, nothing falls through cracks.
Does this proposed vendor contract meet our data retention and cross-border transfer requirements?
Route agents to Corti Models for EU-hosted, OpenAI-compatible inference, or run the Agentic API in the EU or US project environment you already use.
const myAgent = await client.agents.create({ name: "My First Agent", description: "A simple agent to get started with the Corti Agentic Framework", lifecycle: "persistent", connectors: [ { type: "registry", name: "memory-expert" }, ],});A2A is how your application talks to Corti agents: messages, tasks, streaming. MCP is one of five connector types: how an agent calls external tool servers. Different layers, complementary.
It plans work, selects connectors, and returns results. Quick operations come back as messages. Longer workflows become tasks you can monitor, with artifacts for the output.
The Agentic API runs in EU and US environments at api.eu.corti.app and api.us.corti.app. Pair it with Corti Models in the matching region. EU projects get EU sovereign hosting.
Quick operations like classifications come back as messages. Longer workflows that coordinate multiple connectors or wait on downstream systems become tasks you can monitor, with artifacts for the output.
Text parts are for instructions the agent reasons about directly. Data parts carry structured JSON that gets stored in context memory and retrieved when needed. You can use both in one message.
No. Contexts provide strict data isolation. If you need to share information across contexts, pass it explicitly via data parts in your messages.
It analyzes the request and reasons about what connectors are needed. You can guide selection by writing system prompts in the agent configuration. The runtime can chain connectors in sequence or parallel.
Agentic Framework
Create an agent from the API, or talk through connectors and deployment with the team that operates the infrastructure.