Notes on evidence and coding agents
Why agents guess when they only have a ticket, and what changes when the evidence is joined before anyone asks for it.
- Engineering
July 28, 2026
8 minute read
Why your triage agent proposes the wrong fix
A ticket says checkout failed. The click, the failed request, the slow query, the changed row, and the commit sit in five systems that share no key. Why that join has to be stamped as it happens.
- Engineering
July 28, 2026
7 minute read
What is actually in a fix bundle
A recorded session is a directory of plain files: an append only event log, a manifest, and a precomputed evidence index. The layout, and the order an agent walks it in.
- Positioning
July 28, 2026
6 minute read
Session replay is not evidence for an agent
Replay is a video for a person to watch, and it is good at that. A model cannot watch. The difference is the format the evidence arrives in, not the quality of the tool.
- Engineering
July 28, 2026
7 minute read
The bugs that throw no error
Rage clicks, retry storms, slow responses, and abandoned flows are failures with no exception and usually no ticket. Four detectors, four thresholds, and what each one actually catches.
- Walkthrough
July 28, 2026
9 minute read
Give your coding agent bug context in one evening
A walkthrough of the first setup: record a real session, correlate it, and let your agent read the evidence over MCP. MIT licensed packages, no card to start.
- Positioning
July 28, 2026
6 minute read
Tickets are claims. Sessions are evidence.
A support ticket is a secondhand account of a failure. Point a coding agent at one and it will answer confidently from a story rather than from what happened.
Try it on a bug you already have
The capture server and the SDKs are MIT licensed and need no account. One command wires them into your app.