Agents that use your APIs and wait for sign-off
We build agents that carry out multi-step tasks using only the APIs you allow, stop for approval at the steps that matter and record every action.
Who it's for
For companies with repetitive tasks that cut across several systems and still need a person's eye.
What it solves
- Someone opens several systems to handle a request that always follows the same script.
- An AI automation was tried, and nobody could say why it did what it did.
- Giving an agent full access to the systems feels risky, and it is.
- Nobody can say who approved an action that was carried out automatically.
What we do
- Permitted tools
- The agent calls only the APIs opened to it, with its own credentials and the least access needed.
- Approval points
- Payments, messages to customers and changes to records wait for a person to approve or correct them.
- Action limits
- Caps on value, quantity and retries; past a limit, the agent stops and asks for help.
- A record of every step
- What the agent read, proposed and called, and who approved it, kept for audit.
- A workflow that can wait
- The task waits for approval without getting lost; if nobody responds, someone is notified.
- Testing failure scenarios
- Ambiguous requests, missing data and APIs that are down, tested before production.
- An off switch
- One command takes the agent out of service without affecting the rest of the system.
What you get
- An agent in production for the agreed tasks, using only the permitted tools
- A permissions map: what the agent does alone and what needs approval
- An approval queue in the channel your team already uses
- An audit trail for every run
- Failure-scenario tests running on every release
How we do it
Choosing the task
A task with a clear start, end and owner, ideally a repetitive one.
Permissions
We agree with your team what the agent may do and what needs approval.
Supervised pilot
The agent proposes and a person approves every step.
Tuning approvals
Low-risk steps may skip approval if your team so decides, based on the records.
Follow-up
Reviewing the records, refusals and errors, and adjusting permissions.
Technology examples
- Python
- OpenAPI
- JSON Schema
- Temporal
- PostgreSQL
- OpenTelemetry
Related services
FAQ
Does the agent do everything on its own?
No. It carries out the steps your team has allowed and stops at those that need approval. We start with approval on every step; reducing that is your decision, based on the records.
What if the agent does something wrong?
It can happen, because language models make mistakes. That is why high-impact actions wait for approval, limits block anything out of the ordinary and every step is recorded so it can be corrected and, where possible, undone.
Does every system need an API?
It helps a great deal. Without one, file exchange or a new API in front of the system can sometimes work; automating screen clicks is the last resort, because it breaks easily.
Can the agent make decisions about customers?
Decisions that affect people, such as turning down a request or cancelling a contract, stay with a person: the agent prepares and explains. Legal assessment, including under the LGPD, stays with your legal team.
Write to Balkan
Talk to us about your project