AI agent development
Agent workflows that plan, build and check real work, with a person approving each stage.
Start a conversationAgents are useful when they are treated like a disciplined engineering team: clear instructions, small tasks, and someone else checking the work. That is how we use them every day, across a library of more than 40 written agent skills.
We build the same kind of workflow for client teams. You get agents that do real work, and a process that shows what they did and why.
01Deliverables
What you get
- A written specification with acceptance criteria, agreed before any build starts
- Agent workflows scoped to your tasks, with clear permissions and approval points
- Reusable agent skills: written, versioned playbooks your team can read and change
- An independent review step, so no agent marks its own work
- Tests and browser-level checks that run before work is accepted
- A written record of the decisions behind every task
02Process
How we work
One loop for our products and for client work. Every step is written down.
Specify
We write down the outcome, the users, the constraints and the acceptance criteria.
Plan
The work becomes small, testable tasks, each with exact files, interfaces and tests.
Build and review
A scoped agent builds each task test-first. A separate reviewer checks it against the spec.
Verify
Tests and a real browser confirm the result before you accept it.
03Proof
Where we already do this
Our own products and notes, open to read.
Capability
Agentic engineeringSpec-driven agent workflows that plan, build, review and verify software, with a human approving each stage.
Capability
Research and evaluationRecurring, ranked technical research that informs model, tooling and product decisions.
Blog
How we ship software with AI agents without losing controlSpecs, small plans, task-scoped agents, independent review and real checks. How agents do the work while people keep the decisions.
Blog
LLM evals for small teams: a practical starting pointHow to build LLM evals with a small team: start from real failures, use pass/fail checks, calibrate an LLM judge and gate every prompt change.
04FAQ
Questions
- What is AI agent development?
- Designing and building software agents that carry out multi-step work: reading context, using tools, producing output and checking it. We build them as scoped workflows with human approval points, not as open-ended automation.
- How do you keep agents under control?
- Scoped permissions, no destructive action without approval, a separate reviewer for every task and an audit trail. People decide. Agents propose, implement and verify.
- Which tools do you build with?
- Our own workflows run on Claude Code with custom agent skills, subagent orchestration, Playwright and MCP integrations. For client work we choose tools per project and write down why.
- Can you set this up inside our team?
- Yes. We start from a written spec and build the workflow around your repository and your review process, so your team can run and change it after we leave.
Work with us
Tell us what you want built.
A short form. We reply to every message about client work.
Contact us