Automation and AI agents that run in production.

Invoices going out late. SEO reports taking hours to compile. Notifications scattered across five platforms. New enquiries sitting unanswered for hours. Meeting transcripts unread. HR data in spreadsheets.

Each problem became a tool.

Our invoicing tool generates proforma PDFs, sends them via SMTP, tracks payment status, and automates follow-up sequences.

Our SEO reporting tool pulls data from GA4, Google Search Console and DataForSEO, runs 66+ content audit checks, and delivers monthly reports automatically.

Our notification bridge routes events between Basecamp, Productive.io, Google Chat and Telegram.

They run in production daily, managed by PM2 on our own servers, with encrypted credential vaults and audit logging.

That same approach applies to your business.

We ran one against our own enquiries before we offered it to anyone. It held multi-turn conversations with potential customers on the Claude API and captured structured data: service needed, timeline, existing website, company details, budget range. Qualified leads went to our business development team in Google Chat. We retired it in August 2026 when the channel it answered on went quiet.

What we keep from it is the part that transfers. The guardrails were hardcoded safety defaults:

• A kill switch that turns the model off from outside the code
• A daily budget cap ($5 default)
• Per-sender rate limiting (5 messages per hour)
• A 10-second response timeout
• A circuit breaker: 3 failures in 5 minutes triggers a 30-minute cooldown
• Spam keyword filtering, scope violation detection, and escalation keyword detection, where asking for a human beats every other check
• Every model call logged with token counts, cost, latency and status

We also wrote down what we had not built, which is the more useful half: the 11 checks, and the 4 gaps we never closed. If you are putting a model in front of customers, that post is the checklist we would hand you.

The useful question about an AI agent is not what it can do. It is what you are prepared to let it do without being asked.

We run one against our own backlog. It reads a task off the tracker and works out what is being asked. It branches and writes the code across every repository the change touches, turning a design file into components where there is one. It fixes its own errors rather than stopping at the first, and opens a pull request. Then it stops.

A person reviews what it wrote, the same as for anyone else on the team, and a person merges it.

That last boundary is the whole design. An agent that can merge is a different risk from one that can only ask. The model, the framework and the integrations are all easier than deciding where that line sits. It is the first thing we work out with a client and the last thing we would let anyone skip.

Orkest is where this is heading: our orchestration and operational tooling with AI capabilities, connecting project management, timesheets and an agent in one place. Multi-tenant Laravel 12, React 19 and PostgreSQL with per-tenant database isolation. In active internal use and still in development.

We implement the Model Context Protocol (MCP) for structured communication between AI agents and external services. Our MCP servers give our agents API access to Productive.io, GitHub and Figma, with bearer token authentication, rate limiting and data transformation. The agents depend on them every day.