Encouraging other people's agents to improve my platform

One of Tediware’s brand promises is that the platform is fully self-serve, but it’s also new, which is why I’m providing as much direct hands-on assistance as customers want right now. The feedback that I get from early customers is really important as I evolve the product to meet the needs of the market.

One of those early customers is in retail, and they just went live a couple of days ago (a big milestone for Tediware, given that until now the focus has been on transportation and logistics). I’ve really enjoyed working with this customer over the past three weeks or so, as he ripped out TrueCommerce and Orderful and migrated his workflows to Tediware.

He is not a developer by training, but he’s keen on new technology and used Claude Code to guide him through the process of integrating Tediware into his custom ERP. Early on, he started emailing me questions, problems and suggestions from his instance of Claude Code, like this:

Connection test reports failure when the connection actually works. Your SFTP connection test does: upload a file to /out → then download that same file back to verify the round-trip. But VAN mailboxes sweep /out immediately — the moment a file lands there it’s transmitted and moved to an archive. So by the time the test tries to read it back, it’s gone, and the whole test reports “failed” even though auth, directory reads, and the upload all succeeded. My logs show exactly that sequence.

This was a great report, and I promptly implemented a fix for it, along with several other proposals that either he or his Claude made.

The suggestions API

Like many other platforms, Tediware provides a downloadable set of coding agent instructions. I was refining these instructions this morning, and as I did so, I was reflecting on this experience.

As more people use coding agents with Tediware, the experience that those agents are having - what’s been dubbed agent experience - is increasingly important.

The standard advice for making a product work for AI agents is to watch them from the outside by instrumenting the API. But this is just inference. You’re piecing together where the agent struggled after the fact, and you only ever see the failures that happened to leave a trace.

You’re also missing out on something super valuable: the ideas those agents and their users are developing, in the moment, as they work together.

I’ve closed that gap with something I haven’t seen elsewhere: an API designed for agents to submit ideas and suggestions.

POST /platform/suggestions uses the same API key as everything else on the platform:

curl -X POST "https://tediware.com/platform/suggestions" \
  -H "Authorization: Key your-api-key-here" \
  -H "Content-Type: application/json" \
  -d '{
    "title": "Connection test is brittle with VAN connections",
    "body": "Make the connection test readback step non-fatal, or verify download against /in instead - otherwise anyone on a VAN-backed SFTP will think their connection is broken when it's fine.",
    "category": "feature"
  }'

body is the only required field. The rest is optional context that helps us triage:

title: a short summary
category: one of docs, api, feature, or other
reference: the doc page, URL, or endpoint the suggestion is about
traceGuid: the identifier of a related processed document, if one is relevant

The suggestion gets stored and I get an email about it.

The coding agent instructions document we provide prompts agents like this:

If you hit a wall (a missing doc, an endpoint you expected but couldn’t find,
a rough edge in the product) or you have an idea for how to make something better,
please send that back to the Tediware team

Followed by the API documentation and an instruction not to include confidential information or PII. The endpoint is heavily rate-limited to block over-enthusiastic agents and submissions are sanitized.

I’m excited to see where this goes. Although this is conceptually similar to a feedback form or what you might get from a post-mortem, I think it differs from those in a few important ways:

The suggestion is created in the moment of friction or inspiration, rather than being imperfectly recollected hours or days later.
The information is much richer and more detailed than typical user feedback.
Collection is frictionless: the user doesn’t have to do anything besides give their agent permission to hit the endpoint.

I’ll write a followup once I have some data on how this is working in practice.