Agents that do the work and leave a paper trail.
Most products sold as AI agents are chat interfaces. They answer questions, then a person does the actual work. A useful agent updates the CRM record, drafts the refund, files the ticket, chases the overdue invoice. That means giving software write access to systems that matter, which is where most teams stop. They are right to be careful.
We build agents that act inside guardrails. Every action type is granted explicitly, and anything not granted is refused. Risky actions such as refunds over a threshold, emails to customers, or changes to billing sit in an approval queue until a named person clicks approve. Every step lands in an audit log you can read line by line. Model calls run on OpenAI or Anthropic with a hard cost ceiling, so a runaway loop stops itself instead of running up the bill.
Engineers design the permission model and everything touching money or personal data by hand, and read every change before it ships. A narrow single-workflow agent is usually live in 2–4 weeks; agents that act across several systems take 3–8. Either way there is a clickable URL by the end of week one and a written update every day.
Established businesses with a queue that never empties. Support tickets that follow the same dozen patterns. Supplier emails where three fields need copying into an ERP. Invoices to chase at day 30, day 45, day 60. If a person does it today with a checklist, an agent can usually do the first pass and route the exceptions to a human.
Startups building an agent as the product. If the pitch is an AI that does the work for a specific industry, we build the agent plus the parts investors never see: the approval flow, the rate limits, the audit trail, the cost ceiling. Startups can engage every service we offer, not just the four listed under startups.
Teams that tried an agent framework demo that worked in the screencast and fell over on real data. We are happy to start from what you have rather than insisting on a rewrite.
Scoped in writing. Delivered in your accounts.
Explicit action permissions
Each action the agent can take is declared up front; anything not declared is refused, not attempted.
Human approval queues
Risky actions wait for a named person to approve, with the agent's reasoning attached to each request.
Full audit log
Every action records what triggered it, what changed, and when, so any incident can be reconstructed.
Hard cost ceilings
Model spend on OpenAI or Anthropic is capped in code, so a bad loop stops itself instead of running all night.
Retries and dead-letter queues
Failed actions retry with backoff on Inngest; anything that keeps failing lands in a queue a human reviews.
Second-model code review
Generated code is checked by a second model against its requirement, then an engineer reads every change.
Your repo from day one
The repository is created in your GitHub organisation on day one, with full IP assignment.
30-day fix window
Anything we built that breaks in the first 30 days after launch is fixed at no cost.
Scope and fixed quote
A scoping call, with a mutual NDA signed the same day before you share anything. One fixed price in writing within 48 hours, and the number does not change after kickoff.
First workflow live
We pick the narrowest useful workflow and ship it behind an approval queue. You get a live clickable URL by the end of week one and a written update every day.
Widen the remit
As the audit log builds a track record, we add actions and loosen approvals where you decide the record supports it. Permissions and anything touching money stay engineer-designed.
Launch and support
Anything we built that breaks in the first 30 days is fixed at no cost. After that, optional monthly care you can stop at any time.