Centaur: A Multiplayer, Self‑Hosted AI Agent Runtime for Teams

The platform acts as a control plane for agents, coordinating runtime execution, tool access, workflow orchestration, and persistent memory so that agents can perform meaningful work rather than just respond to isolated prompts.

How the Multiplayer Agent Model Works

Most AI assistants today are designed around a single user interacting with a model. Centaur instead treats agents as shared collaborators for a team.

Within a Slack thread or API interaction, the agent can:

• interpret user instructions
• call internal or external tools
• perform multi‑step workflows
• report progress back into the conversation

Because the agent is shared, team members can observe, steer, or extend its work in the same conversation. This collaborative design allows agents to function more like “virtual teammates” embedded in everyday communication tools.

The runtime can also expose internal services as typed tools, letting agents interact with company systems such as databases, APIs, or internal dashboards.

Long‑Running Processes and Durable Workflows

Centaur is designed to support long‑running tasks rather than only short conversational requests.

Agents can:

• run workflows for hours or days
• continue execution across system restarts
• coordinate multiple steps or tools during a task

These capabilities make the platform suitable for automation scenarios that require persistence, monitoring, or iterative progress over time.

Durable execution is a key part of the design: the system records each stage of a request—from user input through runtime execution and final output—so work can resume or be inspected if something fails.

Persistent State and PostgreSQL Storage

Centaur stores operational state in PostgreSQL, enabling durable tracking of agent activity.

According to the project documentation, the database records elements such as:

• the user request
• runtime assignment
• execution events and streamed outputs
• terminal states and final responses

Because these events are persisted, the system can recover from restarts and maintain a history of agent interactions and workflow progress.

Security and Credential Management

A major design goal of Centaur is allowing agents to use real credentials safely.

The system prevents agents from directly accessing long‑lived secrets. Instead, credentials are injected at the network layer during requests, ensuring that the agent never sees raw API keys or tokens.

This architecture helps protect against risks common in agent systems, including:

• prompt‑injection attacks
• malicious instructions in external content
• accidental secret leakage

Centaur’s security model assumes that agents may execute untrusted code or interact with adversarial inputs, so it isolates runtime environments and tightly controls tool access.

How Paradigm Uses Centaur Internally

Before open‑sourcing the system, Paradigm deployed Centaur internally across several teams. The firm reports it uses the platform across a broad set of operational areas, including:

• research and investment analysis
• engineering workflows
• design work
• recruiting and hiring processes
• event coordination
• customer support

According to the company, introducing shared agents across these functions significantly changed how work is organized and executed within the organization.

Extending Centaur With Tools and Workflows

Organizations can customize Centaur by adding internal tools and workflows.

The platform allows teams to expose internal services as structured Python tools and package extensions in overlays that add organization‑specific prompts, workflows, or integrations.

This approach enables companies to build a shared agent ecosystem without rewriting their infrastructure for each automation or assistant.

Centaur’s Role in the Open‑Source Agent Ecosystem

Centaur enters a rapidly expanding ecosystem of agent frameworks and orchestration platforms. Its distinctive focus is on team‑oriented operations inside enterprise infrastructure.

Rather than emphasizing experimental autonomous agents, the platform focuses on practical operational guarantees such as:

• durable memory
• secure credential handling
• workflow orchestration
• collaborative agent interaction

By combining these features with Slack integration and self‑hosting, Centaur targets organizations that want AI agents performing real work inside existing workflows rather than isolated chatbot experiments.

Why Platforms Like Centaur Are Emerging

As companies deploy AI agents in production environments, reliability and security become central concerns. Platforms like Centaur aim to provide the shared infrastructure needed to manage those agents safely at scale.

Instead of building custom agent systems for every application, teams can use a runtime that standardizes execution, storage, credentials, and collaboration. The result is a foundation where AI agents behave less like temporary assistants and more like persistent software systems embedded in everyday work.

Centaur’s open‑source release suggests that this agent runtime layer—similar to how Kubernetes standardized container orchestration—may become a critical piece of the next generation of AI‑powered software stacks.