TRANXACT.AI मदद

Agentic AI: Switch on and off, cost ceilings, the model it will use, tool calls, a fixed code and a local name

Agents can be switched off and on from Next step with a reason; the per-run cost ceiling cannot exceed the per-day one and both are in the company currency alongside the guardrails; an empty Model says which company default applies; runs count Tool calls and Calls succeeded; an agent’s Code is fixed once it is created; and Name in local language is shown to non-English readers.

Agentic AI › Build & Orchestrate › Build › Agents.

Switch off and Switch on

  1. On an agent, choose Next step › Switch off (or Switch on).
  2. Give a Reason — it is recorded against your name in History.
  3. A switched-off agent does not run, by hand or on its triggers, until it is switched on again; its status, versions and history are kept. Switching on a draft does not make it runnable: Publish does.
  • The Agents list and the agent’s header show On or off beside Status (switched on / switched off).
  • Needs Change on Agents; not offered on a retired agent. The Enabled box on Edit does the same.

Cost ceilings

  • Cost ceiling per run (company currency): a run is stopped when it has cost this much. The company guardrail AG-SPEND-RUN (Agentic AI › Guardrails) applies as well; whichever is lower stops the run.
  • Cost ceiling per day (company currency): what this agent’s runs may cost together in a day. The guardrail AG-SPEND-DAY caps all agents together; whichever is reached first stops the run.
  • Empty: only the guardrail applies. A ceiling must be zero or more.
  • A per-run ceiling above the per-day one is refused: “Cost ceiling per run (<50>) cannot be more than the cost ceiling per day (<20>).” One run is part of a day, so it could never be reached.
  • Both are now applied before every step of a run, and the run says which ceiling stopped it.

Model and Engine

  • An agent with no model says “Not set (company default: <model> on <provider>) — choose a model”, or “Not set — choose a model” when the company has no default.
  • The list shows Engine — the provider, or “<provider> (company default)” — and Model used (last run), which reads “No run yet” for an agent that has not run.
  • An agent saved on an engine that is no longer offered says “<engine> (no longer available)” and to choose another engine or leave Engine empty for the company default. Since 21 October Engine is required on a new agent and starts at the company default.

Tool calls on runs

Agent Runs shows Tool calls and Calls succeeded (they replace Calls), counted as the run’s Tool Calls card lists them; the run header reads “N tool calls · M succeeded”. A call left open when a run ends is closed as failed, “Not completed: <reason>”. Agent Runs also shows Agent version and Workflow version, with Agent version used and Workflow version used to open them.

Code is fixed once the agent is created

Edit shows Code locked: it is set when the agent is created and cannot change afterwards, because the audit trail, its runs, versions and findings carry it. To use another code, copy the agent (Copy gives the copy a new code) and switch this one off.

Name in local language

Name in local language (was Local name) is the agent’s name in the company’s local language — Arabic in Saudi Arabia, Hindi or a regional language in India. People using the console in a language other than English see it; English readers see Name. Leave it empty to show Name to everyone. It is written right to left for Arabic by itself.

How to reuse: replace each value in angle brackets with your own amounts, model and provider.

Since 20 October an agent cannot read or change an area whose AI is off, and under suggest-only every change it makes waits for approval. See AI governance: AI Settings per area, AI Decision Reviews, the AI-use report and the AI Disclosure Register.

Since 21 October Engine and Kind of work are required, Switch off and Switch on say what they do with an optional Reason, and Change autonomy asks for a reason only when raising it. See Agentic AI: evaluations by criterion, the 360 cards, Engine and Model, workflow nodes, runs that read every record, and incidents that say what happened.