Introduction

A source is an agent credited with producing an outcome. Adding one is how you say "this agent also does this, and here is how I know".

The rule belongs to the source, not the outcome, because two agents can reach the same result by different routes — and forcing them to share a rule would mean one of them counts wrongly.

Where to find it

AI Governance > Outcomes & ROI > select an outcome > Add agent

Add a source

Step 1: Open the outcome and click Add agent

The picker searches registered agents. Only registered agents can be credited — a discovered but unregistered agent is not available here, which is one more reason registration matters.

Step 2: Define when one of its sessions counts

The panel reads <Agent> — define when one of its sessions counts. Under How it is verified, choose Rule, AI judge or Automation, then set What it is worth — A set amount or An automation prices it.

Step 3: Configure the check

A rule takes conditions; a judge takes a model, a criterion and a pass score threshold; an automation takes a Recorded by automation. See Deterministic or LLM Judge.

Step 4: Read the summary line

Each source summarises as Counts when … or Graded by an AI judge, passes at …. Read it back — it is the rule in plain English and where a wrong condition becomes obvious.

Several agents, one outcome

This is the common case and the reason the model is shaped this way. A ticket might be resolved by a triage agent or by a specialist agent, and both deserve credit when they do it.

1. Each source gets its own rule — The triage agent may be judged on a tool call; the specialist on an evaluator score.

2. Each agent has its own value — Set in its What it is worth card. Use the same set amount on each if a resolved ticket is worth the same whoever resolved it.

3. The Value by score table shows the split — Source, Return, Owner, Total value, Total spend and Net value per agent, with an edit pencil per row and Add agent. Which is how you find out that one agent carries the outcome and another was never really contributing.

Above that table, the outcome page has an Objective card: Total outcomes, Return on spend, Total value, Total spend, Net value and a per-day chart.

Sources need maintenance

This is the failure mode nobody anticipates. A rule reads a signal the session carries — a tool name, an evaluator, a tag. When any of those changes, the rule silently stops matching.

Nothing warns you. The outcome simply reports fewer occurrences, and a falling count looks like falling performance rather than a broken rule.

1. Review sources when an agent changes — Especially a tool rename or an evaluator being removed.

2. Investigate a count that drops to zero — Before concluding the agent stopped working, check that the rule still matches.

3. Watch the Value by score table — One source going quiet while others continue is the clearest signal that a rule has broken.

When a source shows nothing