Dynamiq

Every workflow is a graph of nodes with an input and an output at the ends, and the canvas and the Python SDK compile to the same definition. Analysts drag nodes from a palette grouped by logic, agents, tools and validators. Engineers keep the same graph in git, run it through their CI, and export a canvas workflow as a ZIP when they'd rather extend it in code. The dynamiq CLI handles projects, versions, apps and triggers from a terminal.

You describe the job in a sentence, such as reviewing a loan package against credit policy and routing exceptions to an underwriter, and the generator drafts the nodes, edges and prompts in a side panel. You accept the draft, then swap models, attach tools and add approval steps in the inspector. Nothing is saved until you click Save, so an experiment costs you a minute and no cleanup.

Some decisions shouldn't go to a model at all. A Decision Table turns an eligibility check or pricing grid into typed inputs, a rule grid and a hit policy that a credit lead can read and edit, with CSV import and saved test cases that re-run after every change. A Rules node checks a record against several rules and reports one finding per rule with a severity. Where judgement is needed, the Judgement node answers yes or no, choice and score questions with a confidence score and sends low-confidence cases to a person.

Turn on execution approval for a tool node and the run pauses with a rendered message showing exactly what will execute. A reviewer approves, edits the fields about to be sent, or rejects with feedback the agent reads and acts on. Approval requests arrive over the same WebSocket or Runs API stream the app already uses, so the gate lives inside your own interface rather than a separate console.

With memory enabled, an agent keeps each conversation, scoped by user ID and session ID, and replays recent messages, relevant ones or both at the start of the next run. Long-term memory goes further: the agent gets tools to save and recall durable facts, updates a fact in place when it hears the same thing said differently, and keeps sub-agent history separate so a delegate never reads another agent's memory.

Answers are grounded in your own documents, and each source's permissions are enforced at retrieval, so an agent can't quote a file the person asking isn't allowed to read. Guardrails detect PII and prompt injection before they reach the model or the tool, and every step lands in a trace with the inputs, the model call and the latency. An agent reads those traces and points out the runs that went wrong.