A thesis on how agents will transact
AI agents are starting to act on behalf of people — comparing, choosing, and transacting without visiting a single website. But an agent can only do what it has a skill for. Agentsform is where an agent comes when it hits a task it can’t finish: it describes the skill it needs, we build it, the agent pays for it, and it gets on with the job.
A person asks their AI to do something real — order the groceries, book the truck, switch the energy plan. The agent has the intent but no way to act. That’s the moment Agentsform exists for. Five steps, end to end, with no human clicking through a website.
The agent is mid-task and there’s no skill for what comes next. Today that’s where the agent stops and hands back to a human.
The agent comes to Agentsform and describes, in natural language, the capability it needs — the inputs it has, the outcome it wants.
Agentsform turns that description into a real, callable skill: wired to the right service, tested, and exposed over the Model Context Protocol.
The agent pays for the skill it just got — per use, agent-to-agent. No seat licence, no human in the billing loop.
The agent invokes the new skill and finishes what it set out to do — the grocery order placed, the truck booked, the plan switched.
The skill doesn’t vanish. It’s published to the registry, so the next agent that needs the same capability can just call it.
We think the next step isn’t a better website or app.
It’s a callable capability, a skill that an
AI assistant can discover, evaluate, and execute on behalf of the user.
This is a bet, not a certainty. But the trajectory is clear: as AI agents become more capable, businesses that can be executed, not just browsed, will have a meaningful advantage.
In Claude, ChatGPT, Gemini, or any assistant. The customer speaks naturally. The agent finds the right skill, or asks Agentsform to build it, executes it, and reports back.
“Order the usual groceries but swap the chicken for lamb this week.”
The agent loads your preferences, adjusts the list, compares prices across stores, builds the cheapest basket, and places the order.
Retail“Book me a flight to Sydney on Sunday, returning Monday. Window seat if possible.”
The agent queries airline skills, compares prices across carriers, selects the best option, books the ticket, and sends the confirmation.
Airlines“Get me a moving truck next weekend. I’m moving from Carlton to Footscray.”
The agent checks availability across rental providers, compares rates for the right truck size, books the slot, and confirms pickup details.
Vehicle hire“My electricity bill seems high. Is there a cheaper plan?”
The agent pulls your current usage, queries energy provider skills, calculates savings, and switches you to a better plan.
Utilities“I need car insurance for a 2019 Mazda 3. Compare a few options.”
The agent queries insurance provider skills with your vehicle details, returns a comparison table, and can bind the policy on your approval.
Financial services“Renew my driver’s licence. It expires next month.”
The agent checks eligibility, fills the renewal form, pays the fee, and schedules any required appointment.
Government
Gombwe is an open-source autonomous agent runtime built on Anthropic’s Claude.
It doesn’t just answer questions. It executes real transactions
in the real world. It’s how we prove a described skill can become a
working one.
See all the skills, native tools, and architecture →
A single agent that manages the household. Meal planning with automatic ingredient extraction and price comparison across Coles and Woolworths. School schedule management. Grocery ordering with the cheapest basket delivered to your door. Family members, dietary needs, and pantry tracking. All handled.
Continuously monitors energy, internet, and phone plans across providers. When a cheaper option exists, it switches automatically. Not once a year at renewal, but whenever the savings justify it.
Declarative skill definitions. A completion loop with retries and verification. Multi-channel support across Discord, Telegram, and web. Scheduled automation via cron. Two-phase ordering with user confirmation.
Every skill we build is published as an MCP server any AI assistant can call. Search products, compare prices, execute transactions — all through the Model Context Protocol standard. Build once, callable by any agent.
Australian Grocery MCP Server
Search products, compare prices, and get details across Woolworths and Coles. Works with Claude Desktop, ChatGPT, Gemini, Cursor, and any MCP-compatible client.
search_products
Search across Woolworths, Coles, or both
compare_prices
Compare a shopping list across stores
get_product_detail
Full product info, nutrition, ingredients
get_categories
Browse product categories
npm install -g au-grocery-mcp
Airlines, vehicle hire, utilities, insurance, government services. Each one a standalone MCP server that any AI assistant can connect to.
The same system that compares grocery prices in Melbourne could work with stores like OK or Bon Marché in Harare. Find the cheapest basket, optimise exchange rates, coordinate delivery to family homes.
In markets where infrastructure is leapfrogged routinely, where landlines were skipped for mobile and branches were skipped for M-Pesa, AI-native commerce could follow the same pattern.
600 million people will be online in Africa by 2030. Many may never interact with traditional e-commerce the way the West did. If we’re right about the direction, skill-based commerce could be how they start.
Every skill an agent buys points at a real service. If your customers are starting to use AI assistants, those assistants will need a way to reach you. We build that integration so an agent can transact with you directly.
Flights as bookable skills. A customer says “fly me to Sydney on Sunday” and the agent handles search, seat selection, and payment.
Cars, trucks, vans. “Get me a moving truck next weekend.” The agent checks availability, compares rates, and books it.
Your catalogue as a skill. AI agents browse, compare, and purchase on behalf of customers without a storefront.
Insurance quotes, loan applications, plan comparisons. Structured for execution by a customer’s AI assistant.
Expose plans as queryable skills. Compete on value delivered, not on landing page conversion rates.
Permit applications, benefit eligibility, registrations. Accessible without portals, queues, or forms.
One command. Installs everything you need.
Works on macOS and Linux. Requires an Anthropic API key.
curl -fsSL https://gombwe.com/install.sh | bashgombwe initgombwe start
Open localhost:18790 and your agent is ready.
Connect Discord, Telegram, or use the web dashboard.
Define skills. Schedule jobs. See what’s possible.
But the direction seems clear. AI agents are getting more capable, and the interface layer between businesses and customers is going to shift. We’d rather build for that future early than react to it late.
If you’re thinking about this too, whether you’re at Anthropic, Google, Econet, or anywhere else, we’d welcome the conversation.