ContactDiscuss an opportunity
Services · AI agent development

AI agents: you set the goal, the agent chooses the steps.

Agents that plan their own steps and use your tools to reach a goal, with permissions, approvals and an audit trail around them, so a person stays accountable.

Discuss your projectSee related cases
An agent acts. A person stays accountable.
Plansplit the goalinto steps1Actcall yourtools2Checktest theresult3Approvesign-off onrisky steps4changes requestedretryScoped permissionsLogged actionsCost limitsAn agent acts. A person stays accountable.
We run five agents on the CRM of our own product, Sales Blackbox
1qualification flow tested first, before any broad build
Technologies behind this work
Models & frameworks
OpenAI
Claude
LangChain
Google Gemini
Mistral
Hugging Face
Ollama
CrewAI
MCP
Memory, tools & runtime
PostgreSQL
Qdrant
Redis
Supabase
Temporal
Docker
Python
Node.js
GitHub
What is included

Three areas, from scope to running in production.

An agent is software with a job description. We write it before we build it.

Design the agent

Say what it may do.
  • Goal, scope and permissions
  • Tools it may call, and retrieval over your own knowledge
  • Success measures agreed up front

Keep it controlled

So it can be trusted.
  • Human approval for irreversible steps
  • Logged actions and an audit trail
  • Evaluation of results before and after release

Run it

In your real systems.
  • Integration with CRM, email and internal tools
  • Monitoring and cost limits
  • A named owner for the agent
Typical situations

When teams come to us.

Idea

You want an agent and are not sure it will work

We test the shortest flow first. In our own test, the shorter flow won, and it changed what we built.

Risk

An agent can act, and that scares your security team

We scope permissions, log every action and require a person to approve anything irreversible, and we write that into the plan.

Scale

A prototype works, and now it has to run every day

We add the evaluation set, monitoring and ownership that turn a prototype into something the team depends on.

What you get

Deliverables you will keep.

Concrete artifacts, documented and handed over.

A written scope: what the agent may and may not do
A working agent connected to your tools
An evaluation set and release checks
Approval steps and an audit trail
Monitoring and cost limits
A runbook and a named owner
How to start

Three ways to begin.

Larger work starts with a short paid discovery.

All offers
Build

AI Workflow / Internal Product Build

AI and internal tools around real workflows, not demos.

Custom scope
Sprint

Validation Sprint

A short proof and decision checkpoint before you commit more time, money or roadmap.

Usually 2–3 weeks
Engagement

AI / Data / Cloud Modernization Engagement

Modernize workflows, data, AI systems and cloud architecture.

Custom scope
Proof

Work you can open and check.

See all 11 case studies
Qualify
before human sales
AI sales qualification assistant
A test redirected the product from replacing sales to qualifying leads before them.
4
pilot success criteria
Legal operations AI pilot
Accuracy, review time, adoption and exception handling, defined before the pilot.
Built
for paid use
AI service for property marketing
Durable processing, tested access and accounting boundaries, and a reviewed security audit.
Where this ends

Not quite what you need?

Is the path known in advance?AI workflow automationNeed controls and testing?Quality, security & complianceWant the idea tested before building?Consulting & product design
FAQ

Questions about ai agent development.

How is an agent different from workflow automation?
In workflow automation the path is known: a trigger, fixed steps, and AI doing one hard step. An agent chooses its own steps toward a goal, using the tools you give it. If the path depends on what the system finds along the way, it is an agent. That is more powerful and needs more control, so we design the permissions and the review step first.
How do you govern agents?
Scoped permissions, logged actions, human approval for irreversible steps, and evaluation before and after release. We wrote the longer version in our blog series on AI-agent governance.
Do you run agents in your own product?
Yes. Sales Blackbox runs five agents on its own CRM, so we build agents for ourselves as well as for clients.

Tell us what the agent should do.

Describe the goal, the tools it would use and what must never happen. We reply with a clear next step, usually within 24 hours.

Discuss your projectBook a routing call