Sagewai

Sagewai is an open-source platform designed for building, deploying, and managing autonomous AI agents. It enables users to describe high-level goals, which are then translated into agent designs, executed in production, and iteratively fine-tuned to reduce operational costs over time. The platform is self-hostable, allowing organizations to run it on their own hardware indefinitely, and is licensed under AGPL-3.0 with alternative licensing available upon request.

The system includes several integrated components tailored for developers and organizations seeking to automate complex workflows. The Python SDK supports building agents that interact with a variety of large language models (LLMs), offering features such as tool calling, memory, workflows, and directives. Its directive library allows prompts to be adapted for different model providers, including Anthropic, OpenAI, and Ollama, without requiring changes to application code. Through the Autopilot feature, users can declare desired outcomes, and Sagewai decomposes these goals into agent missions, selects the appropriate tools, and captures execution traces for further analysis and training data generation. The Curator component converts successful agent runs into JSONL files, which can be used to fine-tune models.

Sagewai’s Fleet module manages the dispatch and routing of tasks across multiple agents, workers, and tenants, ensuring tenant-scoped routing, hard isolation, and capability matching. This supports secure, multi-tenant environments and can be integrated into continuous integration pipelines as a regression gate. The Sealed credential management system provides per-CLI workload identity, vault-backed secrets, and audit trails, ensuring that customer secrets are securely scoped, encrypted, and scrubbed after use.

Observability is a core feature, with real-time dashboards powered by Grafana and OpenTelemetry pipelines that display per-project, per-model, and per-token expenditures. These dashboards allow detailed auditing of AI-related costs and system health. The Training Loop enables organizations to capture outputs from models such as Opus or GPT-5, fine-tune their own small language models on a range of hardware options, and deploy them locally via Ollama, supporting a cost-reduction strategy over time.