We experiment with cutting-edge technology and how it can be used in e-commerce and CRM.
E-commerce
See how you can use the latest tech in your industry to set up and manage your e-commerce store.
CRM
See how you can use the latest tech in your industry to set up and manage your customer relationships.
Inspired by Dan Shipper and Every
View Dan Shipper's post on X →
Goal-Driven Seller Management
The seller sets a goal once. Agents plan, act and report every day, and anything that moves money comes back for a yes.
The question
Marketplace sellers juggle listings, stock, ads and pricing across tools built for clicks, not outcomes. What if the seller stated the outcome they want, and agents worked toward it?
What we explored
Architecture
Goal-driven agents sit on top of an existing commerce platform and reach it through MCP or an app marketplace. They get tools, not database access: read, act, ask. Every action is logged.
What a seller really needs
Outcomes tracked against their own targets (sales, margin, returns), simplicity (the one thing to fix today), and autonomy with a say on the big calls.
Four pivots
Autonomous, goal-driven agents. A human in the loop for pricing, policy and inventory. Auto-correction of routine problems. An inventory of every action the agent has taken.
What we found
- Seller work passes the agent-fit test: a clear, measurable goal, many repeatable actions, and a few high-value human decisions.
- Three use cases carry most of the value: close the sales gap (complete attributes, fix titles for search), fill inventory gaps (pause ads on sold-out items, flag reorders), and find what else to sell (spot demand gaps, draft the new listings).
- The line between agent and seller is money. The agent prepares; the seller approves anything that touches pricing or profitability.
What's next
- Run the three use cases against a real seller's data.
- Measure time saved and goal attainment against a control month.
Free to use
The patterns, tool designs and findings here are open. Take them, build on them, and tell us what you learn.
Voice rail for pre-owned luxury
Browse 120 real pre-owned jackets by talking. Ten taps of filtering become one sentence.
The question
Pre-owned luxury is browse-heavy and filter-heavy. Can voice replace faceted filtering, and keep the screen for looking at the product instead of the controls?
What we explored
The experience
Three jackets lead and ten wait quietly behind them. Say “next” to move on, or refine without starting over: “only black”, “under 400”, “cheaper first”. Every active filter shows as a hang-tag chip.
One tool registry
Nine tools cover the whole experience. Voice (Gemini Live with function calling), typing and browser agents (through WebMCP) all call the same registry, so they change the screen in exactly the same way.
Real catalogue
120 real, unsold jackets, 60 men's and 60 women's, captured from Vestiaire Collective's public search on 12 November 2026 and embedded in the page.
What we found
- Seven spoken turns take a shopper from 120 jackets to one €158 leather jacket.
- Refinement, not search, is the heart of it: shoppers narrow down step by step, and the interface has to keep the state visible.
- Building for voice and for agents at the same time is cheap when both go through one set of tools.
What's next
- Run a live voice session and tune the prompt on real speech.
- Swap the static snapshot for live search, and test a second category.
Free to use
The patterns, tool designs and findings here are open. Take them, build on them, and tell us what you learn.
What an agent can do with Amazon's Seller APIs
The APIs already let an agent fix listings, stock and prices. What is missing is the decision layer on top.
The question
How much of an agent-first seller experience can be built on Amazon's Selling Partner API today, and where does the judgement have to come from us?
What we explored
Access
A seller can connect an agent to their own account for free; serving many sellers means a public app with OAuth sign-in. Access is granted role by role.
Map of the APIs
Ten API families cover listings, catalogue, pricing, notifications, orders, inventory, analytics, customer feedback and buyer messaging. Ads sit outside, in a separate API.
Four propositions
Does the small things, flags what needs the seller, shows blind spots, and works toward a goal. Each was tested against the actual endpoints.
What we found
- Small fixes are fully possible: listings can be patched attribute by attribute, previewed before they go live, and watched for new issues.
- Flagging what needs the seller is possible, but the judgement is ours to build: Amazon offers an expected price to win the featured offer, but no price-elasticity score.
- Blind spots (search gaps, review themes) are visible only to brand owners. Goal tracking works from hourly sales and traffic events plus daily metrics.
- The risks: attribute rules change monthly, versions retire steadily, and sellers distrust AI that edits listings without asking.
What's next
- Build the decision layer: price sensitivity from the agent's own tests.
- Prototype the approval flow sellers will actually trust.
Free to use
The patterns, tool designs and findings here are open. Take them, build on them, and tell us what you learn.
Customer, Context, Intelligence: Where CRM Goes Next
Workflows are now easy to rebuild. CRMs that only move records risk becoming dumb pipes.
The question
If anyone can rebuild CRM workflows in a weekend, what is left to compete on?
What we explored
Customer
The customer starts well before they are in your system: one-to-one with known contacts, one-to-many in communities and channels, many-to-many through partners and creators.
Context
Matching a need to a product is stage one. The harder problems are intent (knowing when someone is actually looking) and reach (getting to them at all).
Channel
Email is polluted, social is noisy, and attention is the scarce currency. Increasingly the audience on the other end is an agent doing the evaluation.
What we found
- Workflow is no longer a moat. Acquisition before a lead exists, and outreach that earns attention, are harder and more defensible.
- Every company needs to be a trusted voice that its customers' agents can understand and choose.
- The open question: who becomes the Truecaller of the CRM era, telling you who is real, who is legitimate and who is worth your time?
- Pricing is likely to move from seats and tasks to outcomes.
What's next
- Part two: context and intelligence, and how to build something that can't be vibe-coded in a weekend.
Free to use
The patterns, tool designs and findings here are open. Take them, build on them, and tell us what you learn.
Agent-native agreements
Four principles that make a document platform agent-native, each shown as one end-to-end flow with the metric to watch.
The question
Contracts suit agents: several teams, repetitive jobs, hand-offs between sales, legal and finance, and a clear finish line in a signature. What does an agreements platform look like when people use it from their own agent?
What we explored
Access
Ask your own agent for an NDA; it drafts over MCP, legal approves inside the chat, and it is sent and signed. Metric: requests completed end to end inside the agent.
Write surface
The document renders beside the chat with fields filled and the one change from the template highlighted. Sales, legal and finance each see their own view.
Intelligence and proactivity
Decisions instead of warnings: “liability is uncapped” becomes a one-click fix from the playbook, and exceptions are remembered with reason, scope and expiry. Renewals arrive drafted 30 days before a contract lapses.
What we found
- Public reviews back three of the four principles indirectly, through workflow pain such as editing after sending and expiring contracts.
- No recent reviewer asks for AI or agents by name. The loudest complaints are about billing and support, which none of the principles address.
- Metered agent access would work against the access principle.
What's next
- Test the write surface with a legal reviewer.
- Track time-to-review and redlines per contract.
Free to use
The patterns, tool designs and findings here are open. Take them, build on them, and tell us what you learn.
An independent concept using PandaDoc as the example, not an official PandaDoc product.
Win the agent's choice, then earn the customer
How a document platform grows when the agent, not the person, picks the tool.
The question
Customers increasingly arrive like this: someone asks for a contract, the agent picks the tool, and that tool earns the customer. How do you grow in that world?
What we explored
Lever 1: be the agent's first choice
Agents choose by reading tool descriptions. Write them for the job (agreements, quotes, NDAs that need signing), not for the API. Metric: how often the agent picks you.
Lever 2: first agreements, no sign-up
Let the first one or two agreements go out on the agent's sign-in, and offer to claim an account once they are signed.
Levers 3 and 4: value before asking, pocket-sized plans
A finished draft before any form. Pay for the few agreements you need, then grow into seats.
What we found
- The funnel now starts inside someone else's assistant. Tool descriptions are the new landing page.
- Fuzzy template matching and failed sends that report success quietly lose the agent's trust.
- Sign-up walls are friction an agent can't get past. Defer them.
What's next
- Measure pick rate across assistants.
- Pilot a no-sign-up first send.
Free to use
The patterns, tool designs and findings here are open. Take them, build on them, and tell us what you learn.
An independent concept using PandaDoc as the example, not an official PandaDoc product.