Define an agent the same way you define a workflow — visually, conversationally, or both. Pick a model. Hand it MCP servers and published workflows as tools. Set memory, termination policy, and guardrails.
Claude, GPT, Gemini, open-weights via Bedrock or your own gateway. Swap per environment — Haiku in staging, Sonnet in production.
Any MCP server. Any published Nexus workflow. The agent's tool list updates the moment the underlying server does.
Short-term turns plus retrieval over your knowledge base. Source ACLs travel with the chunk — agents can't surface what the user can't see.
Cost caps. Step limits. Mark sensitive tools approval-required. The agent halts; the operator portal lights up.
Reasoning, tool inputs, tool outputs — all captured per turn. Inspect any decision the agent made.
Reproduce a failed run with the same inputs. Tweak a prompt, replay, compare. Same harness as workflows.
A/B prompts, models, or tool sets across the same input set. Aggregate cost, latency, and outcome diffs.
We'll walk you through defining a research agent against your own MCP servers — in 30 minutes, on your data.