Agents for regulated work,
built like your other software.
Write agents in YAML and code, release them through pull requests and tests, and keep a person on every risky step. Your data protection officer gets clear answers on where data goes.
Hand back-office work to agents, with people on the risky steps.
Examples for a retail bank's operations team. Each agent is a file in code, each risky action waits for a person, and every decision is on record.
At 06:00 the agent groups the night's rejected transfers by cause and proposes a fix for each. Nothing is resubmitted until the payments team approves it.
An advisor asks in the internal portal. The agent answers from your policy documents with section and page, and the advisor decides what to tell the customer.
A complaint arrives by email. The agent sets the reply deadline and drafts an answer from the case history, and the complaints team edits and sends it.
Less to build, less to explain in every review.
Teams that build agents on their own stack maintain servers, queues and test harnesses, then answer the same data protection questions for every new agent. On connic, the answers live in the project.
Servers, queues and retries
Your own stack: Servers, queues and retry code for each agent
On connic: connic builds and runs the agents your team writes in YAML and code, and retries failed model requests
Report what the agents did and what people decided.
Every run and every approval is recorded, so your operations and risk reports draw on the same data.
Cases per agent
Runs per agent and per case type over time, on a dashboard.
Decisions by people
Every approval request with its decision, who made it and when, in connic and in the audit log.
Processing time
How long runs take on each agent's page, on average and for the slowest ones.
Cost per case
Model cost grouped by a value such as the case type, on the usage page.
Clear answers for your data protection review.
Each stop of a run is a setting your team controls. connic/* model calls run in the EU, and a request fails rather than leave it. You contract with connic GmbH in Munich under a data processing agreement.
- Providers you can name
- The model catalog marks EU-native providers and any provider that keeps request data for 30 days. Managed inference isn't used to train models.
- Core systems stay in your network
- A bridge inside your network opens only outbound connections, and an allow list limits which hosts it can reach.
- Less personal data in run records
- Redact fields such as IBANs from stored runs and shorten how long each environment keeps them.
- Evidence on request
- The audit log records changes and approval decisions. On enterprise, AI Governance exports evidence snapshots for your auditors.
Every change goes through review and tests.
The dispute agent from the top of this page is one YAML file with a test suite next to it. A deploy goes live only when every case passes.
- Checks on pull requests
- On GitHub and GitLab, the suite runs on each pull request and reports
connic/pr-testsas a commit status, so the merge can wait until it passes. - A gate on every deploy
- A push to the branch mapped to Production builds the deployment and runs every case. Until all of them pass, the current version keeps serving.
- Rollback without a rebuild
- Activate an earlier deployment, and it serves again with the connectors it declared.
version: "1.0"
name: dispute-intake
description: "Turns dispute emails into cases"
model: connic/mistral-large-2512
fallback_model: connic/mistral-medium-3.5-128b
system_prompt: |
Read the customer's email and find the disputed card payments.
Propose a provisional credit only when policy D-4 allows it.
tools:
- core_banking.find_card_payments
- cases.open_dispute
- payments.provisional_credit
output_schema: dispute_case
approval:
tools:
- payments.provisional_credit: param.amount_eur > 250version: "1.0"
tests:
- name: double_charge_waits_for_a_person
payload: '{"subject": "Charged twice for my hotel booking"}'
mocks: core_banking
approval_decisions:
- tool: payments.provisional_credit
decision: approve
expected_tool_calls:
- payments.provisional_credit: invocations == 1
expected_result: output.case_type == "duplicate_charge"