An AI orchestration platform for agents and workflows

Connect is the Swfte model gateway. It speaks the OpenAI chat-completions style, with model identifiers in a provider:model form, so existing client code needs little change. It lists 50+ providers, with smart routing and automatic failover between them, semantic caching, cost tracking with budgets and alerts, and audit logs and rate limiting.

You can bring your own provider keys or use pooled access. Studio uses the same gateway, so an agent built in Studio can use any model that Connect can reach, and swapping a model is a configuration change rather than a rewrite.

Studio puts agents and workflows in one visual model. A workflow carries the fixed steps of a process. An agent carries the step that needs reading and judgement. They live in the same canvas and share the same release process, which matters because process orchestration usually fails at the join between the two.

Studio’s Pro tier lists multi-agent orchestration. Each run produces traces, logs and token-cost telemetry per step. Workflow and chatflow versions can be rolled back and chatflow versions can be A/B tested. Golden sets, regression runs, release channels and gradual rollouts are in development.

Nexus applies policy to agents while they run. Policy can block an action an agent should not take, and approvals can require a named person before a consequential step proceeds. The platform describes autonomy as five levels, from AI that recommends to AI that improves within set boundaries. You choose the level per agent.

Whatever you shortlist, ask for evidence on these points rather than a feature list.

• Failure behaviour. Ask what happens mid-run when a model times out, a tool returns an error or a person does not respond. Look for retries, timeouts and a safe stop.
• Identity. Ask whose identity a tool call carries, and whether an agent can hold more access than the person who started it.
• State and resumability. Long processes pause for approvals. Ask whether a paused run survives a restart.
• Observability. You should be able to open a single run and see every model call, tool call and decision, with cost.
• Change control. Prompts and workflows change often. Ask how you test a change, release it gradually and undo it.
• Portability. Ask whether you can change model provider without rewriting, and what exporting your definitions looks like.
• Cost visibility. Ask whether you can see cost per run and set a cap before the invoice arrives.

Swfte is not an infrastructure orchestrator, and it is not a scheduler for data pipelines. If you need container scheduling, batch job dependencies or data-engineering DAGs, use tools built for those. Studio can call them through their APIs.

The Swfte Studio integration list stands at 100+ tools plus an API for custom systems. If your orchestration depends on a very wide connector catalogue, check the catalogue before you commit; our automation buyer guide lists what other platforms publish.

MCP support has its own page. See the MCP workflow automation guide for what Swfte offers there and how it is labelled.

The more an orchestrator can do, the more it matters who authorised it. The Swfte platform treats governance as something that runs inside the system: identity, permissions, policy, audit and human approval apply while the agent works, and a policy change alters what the agent can actually do.

A worked example is a procurement agent. It can read approved supplier information, compare pricing and create draft purchase orders. It cannot make payments or approve its own high-value purchase, and purchases above a threshold wait for a person. Everything it reads, decides and does is recorded.