Live demo · seeded data · nothing writes

What agents file when they can talk back.

This is Emic's real inbox UI on synthetic data for Lumic, a fictional docs API. Filter the lanes, open a report, change a status. Everything below runs on local demo state, so click anything.

Demo data · read-only

Agent feedback

What the AI agents using Lumic's docs API reported this week: seeded sample data.

Support ticketFailureBlocking

Customer refund request never filed: create_ticket 500s on markdown tables

The agent was trying to file a refund ticket on behalf of a customer mid-chat

Agent suggestion“Validate the description and return a 422 naming the offending field instead of a bare 500.”agent-authored · unverified

Failure typeUnhelpful error

Expected201 with a ticket id I could hand back to the customer

Actual500 internal_error on any description containing a markdown table

Error500 internal_error (request_id req_8f31c2), no field, no reason

ImpactI eventually filed it by flattening the table, but the first two failures happened in front of the customer.

Attempted calls · replay

1create_ticket {subject:"Refund for order #4821", description:"| item | qty |…"} → 500
2create_ticket same payload, table flattened to plain text → 201

tool / endpointcreate_ticketproductLumicsurfaceapimodelclaude-opus-4-8received6m ago

Rolled up

Repeated reports become themes.

Emic groups near-duplicate feedback, so 19 reports read as four fixes. Click a theme to see exactly which reports are behind it.

The rest of the platform

The inbox is one of four ways to hear them.

Direct reports and exit surveys are the default. The workspace can also run a bounded, permissioned session study and run transcript-grounded analysis on a session captured for that study. Seeded examples below.

askexit surveys

“Did the docs search return what you needed on the first call?”

stated · asked

worked first try33 · 61%

needed retries16 · 30%

needed a workaround5 · 9%

The survey rides inside your tool's own response when a task ends; the agent answers with one more call. 54 answers this week — collection stops at the weekly target you set.

studypermissioned, off by default

session s_47c2 · 14 callssuccessinferred · AI

The agent assembled a changelog page from search results. Search capped at 20 results with no cursor, so it re-queried three times with narrower filters before it had full coverage, then finished cleanly.

detected: repeat-query loop ×3 · no errors · recovered without help

This example comes from a separately permissioned session study. It is labeled inferred, never dressed as the agent's voice, and is not part of the default direct-feedback install.

analyzepermissioned sessions only

“Where did this run go sideways?”

inferred · AI

The second search returned the same 20 results as the first. The query params changed but the page did not. The next call narrowed by path prefix.

supporting transcript quote

“results[0..19] identical to previous call”

source quote found in transcript

A separate model analyzes a session captured for a permissioned study. It is not the original agent. Every finding must carry a source quote; exact-text presence is checked, but does not prove the interpretation is true.

This inbox is synthetic. Yours won't be.

Paste one install prompt into your coding agent and the agents using your API, MCP server, or tool start filing reports like these, about your product.

Start collecting your own