Human checkpoint
Trigger, then the agent reads and decides, calls your tools, waits for approval, acts, and logs it.
See the checkpointAI Agent Development
AI agents that do real work, with people in control
Agents that handle multi-step work across your tools, with approval points wherever judgement matters.
12+ years shipping web and mobile apps for clients in the US, UK, Australia and India
What changes
Work that runs itself, safely.
The agent moves the work. A person still decides at the approval point.
01
Multi-step work moves across your tools
02
A person approves wherever judgement matters
03
Every action is logged
Is this you?
When this is the right service
Use this when a chatbot reply is not the job. The job is a sequence of steps.
- The work has several steps across more than one tool.
- A reply in a chat window would not finish the task.
- Someone must approve before certain actions.
- You want a pilot on one workflow, not a new platform.
Good fit
- A workflow you can map.
- Tools the agent can call.
- Clear places where a person should decide.
Not a fit
- An open-ended chat with no tools and no owner.
- A process that is not defined yet.
- A decision that must be right with no review.
Services
What we build
One agent, one workflow, with the guardrails written down before the pilot.
Workflow map
The steps, the tools, and the points where a person must approve.
Guardrails
What the agent may do alone, and what waits. Set before the pilot is built.
Pilot agent
A fixed-scope agent on that one workflow. The price is quoted after the map.
Logs
A record of the actions it took, so a wrong step can be seen.
Tool connections
Calls into the systems you already use. Which systems is decided when we map the workflow.
Handover or ongoing care
Documentation so your team can run it, or Continuous Product Engineering if you want us to stay.
How an agent works
Every agent has a human checkpoint
The approval point is a person. The agent does not take that action on its own.
1
Trigger
A form, a message, a time, or a person starts the work.
2
Agent reads and decides
It reads the case and chooses the next step within the guardrails.
3
Calls your tools
It uses the systems the workflow already runs on.
4
Approval point
A person decides wherever judgement matters. The agent does not pass this alone.
5
Action
It takes the step you allowed, and only that step.
6
Logged
The action is recorded so you can see what it did.
How it works
From a map to a monitored agent
Five steps. The pilot stays on one workflow until you approve a wider scope.
- 01
Map the workflow
We write the steps, the tools, and where a person must approve.
You get: that map, before any agent is built.
- 02
Set guardrails and approval points
We mark what it may do alone and what waits.
You get: those rules in writing, and a fixed quote for the pilot.
- 03
Build the pilot
We build the agent for the mapped workflow only.
You get: a pilot you can try on that workflow.
- 04
Test on real cases
We run real cases, including ones it should refuse or get wrong.
You get: the results of those cases, not only the ones it handled well.
- 05
Launch and monitor
We launch the scoped agent and watch the log.
You get: the agent live, with actions logged.
Built for production
How we make it dependable
An agent is production software. These are the checks around the pilot.
Testing
Real cases, including wrong ones, before launch.
Code review
A senior engineer reviews every change, including AI-generated code.
Security checks
What the agent can reach, and which actions need a person.
Monitoring
The action log, so a bad run is visible after launch.
Documentation
The map, the guardrails, and how to stop a step.
Handover
You own the code and the accounts. We hand over how to run the agent.
Technology
Tools we use, chosen per project
The agent calls your tools. The model is chosen during scoping. We do not name one here.
Proof
Work in this area
Client names stay private. Studies are published on Work when they are written.
Ways to work together
How to work with us
Start with a fixed-scope pilot. A wider rollout is a new scope.
Recommended
Fixed-scope pilot
Best for one mapped workflow with approval points.
How it works We map the workflow, set the guardrails, and quote the pilot. No price is listed on this page.
FAQ
Common questions
Answers from the map, the guardrails and the pilot. Tools and the model are named during scoping.
A chatbot replies in a conversation. An agent handles multi-step work across your tools. It stops at an approval point wherever judgement matters, then acts, and the action is logged.
The approval point is there so a person can stop an action that needs judgement. We test the pilot on real cases, including cases it gets wrong, before launch. Actions are logged so you can see what it did.
The tools the workflow already uses. We name them when we map the workflow. There is no fixed tool list on this page.
Before the pilot we set guardrails and approval points: what the agent may do alone, and what waits for a person. The pilot is one workflow. The price is quoted after that map.
After launch we monitor the agent. Ongoing changes can stay with us through Continuous Product Engineering, or we hand it over with documentation. Which of those you want is in the scope.