Write Your Own Agents

Agent definitions in MyBenAI are rows in a database, not code. You can edit the ones that ship and write new ones from the settings screen — no rebuild, no scripting language, no developer mode.

What defines an agent

• A system prompt — the job, in your words.
• A tool allowlist — exactly which of the tools it may call, including your own connectors.
• A step budget — how many tool-and-think rounds before it must answer.
• A temperature — low for extraction, higher for ideas.
• An output schema, optionally.

The schema is the part that makes it work

Asking a 1B or 2B model to "reply in JSON" is a coin flip. MyBenAI instead compiles your schema into a grammar and constrains decoding to it, so the model is structurally unable to emit anything that does not match. That is what makes a small local model usable as a component rather than a chat partner: an Extractor that returns a valid object every single time is worth more than a larger model that returns prose two times in ten.

The design principle underneath

The model is given judgement; the app keeps the books. An agent decides what to search for and what a page means. It does not decide how many rounds it gets, how the citations are numbered, whether a source is a duplicate, or when to stop — all of which are handled in ordinary tested code. Assume a small model and design each step so it cannot be wrong about anything that matters.

Watching them work

Every run is recorded with its steps, its tool calls, its sources and what it cost, so when an answer is wrong you can see where it went wrong instead of guessing.