The tightest loop in agent development is: change the prompt, send a message, read the answer, repeat. Botmanor makes that loop native — and makes every iteration auditable.
Open chat against any agent and send a message. Behind the scenes, Botmanor resolves the agent's model binding to a provider, resolves that provider's vaulted credentials, and calls the LLM. Nothing is mocked: the reply you read in the test chat went through the same execution path your production traffic will.
Every send produces an execution record — status, input, output, and metrics including input/output token counts and cost. Both sides of the exchange are stored in the platform conversations service. When an answer surprises you, you do not guess: you open the execution, see exactly which model ran with what input, and what it cost.
This is what makes prompt iteration honest. Pair the chat loop with prompt templates — versionable prompts with variable placeholders — and you can render a candidate prompt, bind it to the agent, chat-test it, and compare execution metrics across iterations before publishing a new agent version.
Credential resolution is layered for real-world setups: a workflow-injected key, a workspace integration connection, or a stored provider credential — whichever your configuration provides. The test chat is single-turn today (system prompt plus the current message), with conversation-history replay in the runtime roadmap; threads are already persisted, so the history is there waiting.
Do it yourself
Chat-test any agent so each message resolves credentials, calls the real model, and writes an execution record — then open that record to see exactly which model ran, with what tokens and cost.
From the Botmanor sidebar open Chat and pick the agent you want to test.
You should see: A chat surface opens bound to your chosen agent, ready for a message.
Ready to make this your story?
