Agentic AI: evaluations by criterion, the 360 cards, Engine and Model, workflow nodes, runs that read every record, and incidents that say what happened

New Evaluation lists only the chosen agent’s runs and grades a Criterion out of 100; an agent’s 360 view has Evaluations of this agent and Evaluations given by this agent, and every card says what it holds; Kind of work and Engine are required; Change autonomy needs a reason only when raising it; New Node picks the approver role and shows the fields of its kind; runs read every page of a list, count failed calls apart from refused ones and say why they failed; names are shown only to people allowed to open the records; incidents are numbered and say their real cause; ASK knows which form is open.

Agentic AI › Build & Orchestrate › Build › Agents, and the agent’s 360 view.

New Evaluation

• Agent: the agent whose work is graded. Run (optional) lists only that agent’s runs, each with its status; an agent with no runs says “No runs for this agent yet.” Empty grades the agent in general. An evaluation whose run was made by another agent is refused, on create and on edit.
• Criterion (was On): what the grade is for, e.g. <overall>, <accuracy>, <tone>; blank is overall. It is not a date — the date is recorded when the evaluation is saved and is listed as When.
• Score: out of 100 — 0 the worst, 100 the best, e.g. <80>. Optional: Verdict alone is enough.
• Verdict: good (right and useful), acceptable (usable with small fixes), poor (needed real rework), wrong (incorrect) or harmful (would have caused damage if acted on). Kept beside Score; neither is worked out from the other.
• Graded by: A person — you or a colleague; Another agent — an agent acting as grader; What happened next — the result of its work, e.g. a proposal accepted and still standing; A rule — an automatic check.
• Graded by agent is shown, and required, only for Another agent: usually a reviewer agent set up to check other agents’ work.
• Comment says why, in a sentence. Outcome and Value are for What happened next (Accept it and Reject it fill in Outcome).

The agent’s 360 view

• Two Evaluations cards: Evaluations of this agent (grades it received for its work) and Evaluations given by this agent (grades it gave other agents’ work, as the grader). + Add on Evaluations given by this agent opens New Evaluation at Graded by Another agent.
• Every card carries a one-line hint saying what it holds — guardrails, knowledge sources, triggers, tools, versions, artifacts, incidents, run steps, runs, tool calls, approvals, memories and workflow nodes — or an (i) with the related screen’s first sentence.
• Records say what they are instead of “record”: an incident by its number, an evaluation as “<RUN-000003> · <accuracy> · <good>”, a memory by what it remembers (the first sixty characters), and tool calls, nodes, edges, versions and discovered tools by name. Each record page is headed with the same title its card shows.

New Agent and New Workflow

• Kind of work is required. It starts at Operations; choose the one that fits. A blank one is refused, not silently saved as operations.
• Engine is required and starts at the company default engine (Platform › AI Providers). An engine with no key for this company is greyed with the reason and cannot be chosen. An agent created without an Engine is saved with the company default.
• Model lists the models of the chosen Engine and is greyed until one is chosen; a model with no price on file shows its run cost as unknown. A Model is still required before the agent can be published or run.
• Every field of the Agentic AI forms has a hint: Temperature, Max tokens, the tool-call and time limits, memory, guardrail and trigger fields among them.
• A required field filled with spaces counts as empty. What stopped the press is shown under the field, and the form scrolls to it.

Switch on, Switch off and Change autonomy

• Switch off: it stops running — by hand and on its triggers — until switched on again; status, versions and history are kept.
• Switch on: it runs again — by hand and on its triggers. A draft still needs Publish before it can run.
• Reason is optional on both and says so: “Optional. Recorded against your name in History.”
• Change autonomy: Reason is starred, and required, only when you raise the agent’s autonomy. Without one the form says, under Reason, “Raising an agent’s autonomy needs a reason.” Lowering it needs none.

New Node in a workflow

• A line at the top says what a node is: one step of the workflow. Kind decides the fields shown — agent: Agent and Instruction; tool: Tool and Arguments; decision: Expression; loop: Loop over and Most iterations; human: Approver role, Instruction (the question put to the approver) and Answer within (hours); subflow: Sub-workflow and Arguments.
• Approver role is picked from the company’s own roles (Platform › Access › Roles) and is required for a human (approval) step; the server checks it.
• Arguments must be JSON, e.g. {"employee_id": "{{input.employee_id}}"}; it is checked when you save.
• Retries, Time limit (seconds) and On error (inherit, stop, continue, retry, route) say what happens when the step fails.

Corre

• A list tool reads every page of a list: the agent sees whole records, the total count and where the next page starts (100 a page by default, up to 200), so a “check all employees” run checks them all. Findings stay tied to their record even when a tool result is very long.
• An e-mail column is judged by the platform’s own format check, not by the model guessing at a hidden address.
• Failed calls are counted apart from refused ones (a refusal is a guardrail or a person saying no). A run that fails says the cause in its summary, and every step left open is closed when the run ends.
• A failed run, and one halted by a guardrail, counts as a failure on its agent or workflow.
• Names in a run’s summary, output and steps are shown only to people allowed to open every list the run read; anybody else sees “(name withheld)”. Nothing unmasked is stored or sent to a model.

Incidentes

• Every incident is numbered from the AIN series, e.g. <AIN-000004>, including the two segregation-of-duties incidents (an owner approving their own agent; a worker reviewing their own delegated result).
• The title says the real cause with a severity that fits: “The model could not be reached”, “The provider refused the request: the API key was not accepted”, “The provider refused the request: too many requests, or the quota is used up”, “The model wrote a tool call the AI engine refused”, “The run stopped: the database was unavailable or busy” or “The run stopped on an internal error”. Older incidents recorded as unreachable where the provider had in fact answered are restated, with a note in Action taken.
• The Incidents list no longer totals Open (hours) in its footer.

ASK and the AI service

• ASK is told which form is open — its title, the screen it makes a record of, and its fields with their hints and choices (never what you typed). Ask “what is this form about” or “explain the fields” and it answers from that form alone.
• A field the server refuses is said once, under the field, and the form puts the cursor there; the red box at the top is kept for refusals that name no field on the form.
• When the AI provider is down, ASK says “The AI service (<Groq>) is unavailable at the moment (it answered <503>). Try again in a minute, or add a second provider under Platform › AI providers so the assistant can use that one meanwhile.” No defect ticket is raised for the provider’s outage. Short provider errors are retried once first.