Cayu

Cayu is an open-source Python framework for building domain-specific agents whose work may continue for hours or days. It brings the agent harness and the execution runtime into one project, so a team can define the context, tools, policies, memory, business rules, and verification that surround a model. The point isn't to provide a general chat interface. It's to give developers a structured foundation for an agent that performs a bounded job in a real domain and keeps working when the path to completion involves many decisions.

That long-horizon focus addresses a problem that appears after a promising agent prototype meets production. A one-shot prompt can return an answer, but operational work has state, interruptions, external effects, budgets, approvals, and uncertain outcomes. Cayu records sessions, events, and checkpoints in durable storage so progress can survive a stopped worker. A replacement worker can inspect the recorded state and continue, while an action whose effect is uncertain can be reconciled against outside evidence or sent to a person instead of being silently repeated. This makes waiting and recovery part of the workflow rather than exceptional cases bolted on later.

Developers start in ordinary Python. The project generator creates places for prompts, tools, policies, memory, domain code, tests, and evaluations, along with a basic model-only agent to extend. Teams can connect Python functions and business systems, choose a model provider, configure storage, and define the completion checks that decide when work is actually done. Cayu supports OpenAI, Anthropic, Amazon Bedrock, Google Vertex AI, OpenAI-compatible services, and custom provider contracts. Its components are replaceable, so model access, context assembly, knowledge retrieval, storage, and execution don't have to be tied to one vendor.

The integrated runtime covers the less visible machinery behind a working agent. It can keep durable session history, coordinate tasks and workers, pause for human approval, enforce tool policies, reserve and settle model budgets, and track usage through structured events and OpenTelemetry. It also includes an operator dashboard for inspecting sessions, workflows, usage, and evaluations. Tool effects and permissions are represented separately, and approval can be scoped to a particular requested action. Cayu can also place credentials behind configured proxies so an authorized tool can act without receiving the underlying secret, though native Python tools still need an execution boundary appropriate to the code. Recorded sessions can be resumed or forked, while tasks use dispatch, worker leases, and heartbeats to coordinate work beyond one request. Artifacts and approvals can persist across restarts, and a stale worker can be fenced from work it no longer owns. These details help when an agent is a continuing process rather than a single model call.

Cayu fits engineering teams building agents for document processing, finance, accounting, enterprise operations, or other domains where rules and evidence matter as much as model output. It is most useful when a job spans thousands of steps, waits for a reviewer, touches real systems, or must stay within a cost or authority limit. The framework expects the team to supply the actual domain behavior and interface. That makes it a better match for developers creating an agent product than for someone looking for a finished assistant that works without coding.

Its clearest distinction is the attention given to what happens beyond the happy path. Session forks, recovery-aware task execution, durable artifacts, trajectory evaluations, cost attribution, and checks around ambiguous external effects all live in the framework's model of the work. The project also makes its boundaries explicit. Local runners are for trusted code, tool permissions don't create an operating-system sandbox, and evaluations only establish the criteria a team defines. Recorded model usage and estimated cost are treated as separate measures, since an estimate depends on the team's own price book and the usage data a provider returns. Those caveats help set expectations for teams that need to reason carefully about authority and reliability.

The core framework is available under the Apache 2.0 license and installs as a Python package. It requires Python 3.11 or newer and can run locally, on a team's cloud account, in a private network, or on premises. Cayu Cloud is optional and provides a shared workspace for building, deploying, operating, and improving agents, but the public site doesn't publish self-service Cloud prices. The open-source package is therefore the clearest route for evaluating Cayu, while teams that need the managed workspace or enterprise help will need to contact the company.