Loan orchestrator
This page is the agent's profile: how it works, who holds each role, how it is doing, where Auga sees it, and its history.
Check that every role has a person.
Portrait
The picture shows one run at a time, from the same file as the latest Agent Health report. Choose a run, then play it.
Every step ran once, and the policy approved the loan. Of the 400 runs in the report, 169 look like this.
- What happened
- Seven steps in 3.3 seconds. Approved.
- Within the limit?
- Yes. Debt-to-income was 0.20, and the limit is 0.45.
- What it used
- Three tools and no model.
The risk agent asked a language model for help before it gave its score. The loan was approved.
- What happened
- Seven steps in 3.8 seconds. Approved.
- Within the limit?
- Yes. Debt-to-income was 0.22.
- What it used
- model-a-large: 615 tokens in and 101 out. Two tools.
Debt payments took 52% of the income, above the limit of 45%. The policy denied the loan with reason R02, as the rule says.
- What happened
- Seven steps in 3.1 seconds. Denied, R02.
- Within the limit?
- No. Debt-to-income was 0.52, above 0.45.
- What it used
- Four tools and no model. The applicant was told.
Reading the documents failed after its retries, but the run went on and the loan was approved. The identity check was slow, so its dip is the deepest.
- What happened
- Seven steps in 4.7 seconds. Approved.
- Within the limit?
- Yes. Debt-to-income was 0.25.
- What it used
- model-a-large: 568 tokens in and 116 out. One tool failed.
Bureau feed 3.3 reported an income of 43 instead of 43,000. The income check then used the bureau's value, so debt-to-income became 270 and the loan was denied. See the proposal to return the bureau feed to 3.2
- What happened
- Seven steps in 3.5 seconds. Denied, R02.
- Within the limit?
- No. Debt-to-income was 270.5, because the income was read as 43 SEK.
- What it used
- Three tools and no model.
Runs the application through the steps and owns the final decision. Choose a point in the picture to read about it.
- Each point is an agent. Loan orchestrator, this agent, is in the middle and calls the others in order.
- The floor sinks where a step took time. The deeper the dip, the slower the step.
- The ring is the debt-to-income rule. It turns red when a run passes the limit.
- A red peak is a step that failed, or a value above its limit. The further above, the higher the peak.
- A column is a call to a language model. Its height shows the number of tokens.
- Drag the picture with the mouse, or with two fingers, to turn it.
Health
UNHEALTHYThe latest report covers 400 runs of the workflow that this agent starts, on 1 Nov 2026 from 06:00 to 17:58 UTC. Its status is UNHEALTHY.
The strongest change was the bureau feed, from version 3.2 to 3.3. After the change, every run on 3.3 was denied. Of the runs that stayed on 3.2 at the same time, 16% were denied.
The report shows that the change and the denials came together. It does not show that the change caused them. See the proposal to return the bureau feed to 3.2
| This agent | What the report found | What it means |
|---|---|---|
| Response time | 3,586 ms typical. 99% of the calls took less than 5,112 ms. | No alarm. The typical time stayed at its normal level. |
| Telemetry | 94% of its steps carry the application id and the decision. | The other 6% cannot be linked to an application. The technical owner can add the two fields. |
The report covers all 8 agents in the workflow. The status and the strip are for the whole workflow, and the table is for this agent only.
Roles now
| Role | Who | What the role means |
|---|---|---|
| Deployer | Acme Lending | The company that uses the agent. The EU AI Act calls it the deployer. |
| Accountable person | Maria Svensson, Head of Credit | This person answers for what the agent does, and decides about its risks and incidents. |
| Technical owner | Jonas Persson, Platform team | This person keeps the agent running and makes the changes. |
| Human oversight | The credit review team, in Coach | The team reviews the decisions that the policy sends to a person. |
| Model provider | Not set | The telemetry does not name the provider of the model. Add it here. |
Auga records who holds each role. Who is legally accountable is a decision for your company. Each change of owner is signed into the ledger.
Purpose
The agent receives loan applications, asks the other agents for checks, decides, and writes to the applicant.
It belongs to the AI system Loan decisions. Your company classified that system as a high-risk AI system, because it assesses the creditworthiness of people. The EU AI Act lists this in Annex III, point 5(b).
Classified by Maria Svensson on 20 Oct 2026.
Sources
| Source | Kind | Counts for the register | First seen | Last seen |
|---|---|---|---|---|
| prod-eu | OpenTelemetry Collector | Yes, register source | 12 Oct 2026 | 4 Nov 2026 |
| staging | OpenTelemetry Collector | Yes, register source | 2 Oct 2026 | 30 Oct 2026 |
| credit-nov | Uploaded bundle, simulated | No, analysis data | 4 Nov 2026 | 4 Nov 2026 |
Auga knows the agent by the same id in every source: auga.agent.id = loan_orchestrator. The file credit-nov is linked to the agent, but it does not count for the register. .
History
| When | What happened | Who |
|---|---|---|
| 4 Nov 2026 17:26 | Auga opened incident INC-2. | Auga |
| 4 Nov 2026 17:24 | A person marked a letter to an applicant as harmful. | Sam K. |
| 28 Oct 2026 | The stage in External communication was upgraded to Supervised. | Auga |
| 20 Oct 2026 | Maria Svensson became the accountable person. | Ava L. |
| 20 Oct 2026 | The AI system Loan decisions was classified as high-risk. | Maria Svensson |
| 12 Oct 2026 | Auga saw the agent in prod-eu. | Auga |
| 2 Oct 2026 | Auga saw the agent for the first time, in staging. | Auga |
Open
This page is a sketch. Nothing on it is built, and all names and figures are invented.
How the register works
The register lists the agents that run in your organisation. An agent enters the register in two ways: a person declares it, or Auga sees it in a register source.
Register sources and analysis data
- Register sources are the OpenTelemetry Collectors that your organisation uses in production and staging. An admin marks which Collectors count.
- Analysis data is a file that someone uploads, or a bundle in Bundle Analysis. Analysis data never creates an agent in the register. It can only be linked to an agent that is already there.
- A simulated bundle, for example from Rootplant, is never an agent in the register.
Declared and seen
- Declared and seen. A person registered the agent, and Auga sees it in a register source.
- Declared, not seen. A person registered the agent, but it sends no telemetry. It may not be instrumented yet, or it may be retired or set up wrongly.
- Seen, not declared. Auga sees the agent in a register source, but nobody registered it. This can be an agent that nobody governs.
The words for a change of stage
Upgraded and downgraded are the words that credit rating agencies use when they change a rating. Unchanged means that the stage stayed the same. Held lower means that a person keeps the stage below what the evidence allows.
There is no industry standard for the levels of autonomy of AI agents, and the frameworks differ from each other. The rating words are well known to people who work with risk.
Sources
- Fitch Ratings: Rating Definitions. Ratings can be upgraded, downgraded, affirmed or placed on Rating Watch.
- Moody's Rating Terms. Moody's says confirmed where others say affirmed.
- S&P and Moody's rating descriptions, in a filing to the SEC.
- Levels of Autonomy for AI Agents, by Feng, McDonald and Zhang. The most cited research on levels for AI agents.
- AI agent autonomy levels, a framework from practice. It is a blog post, not a standard.
What the register keeps
The register keeps the agents, their roles and their history. The runs stay in the data, for as long as your retention rules say. A retired agent stays in the register with the status Retired, so that an auditor can see what existed and who was accountable. The register covers only your own organisation.