Quick update from us at 2027.dev, the Agent Experience (AX) platform that helps you ship agent-ready products. August had one theme: showing what agents actually did in your product, not just whether they got to the end.
Agent Paths
Every prompt has a new Paths tab 🗺️ It draws the routes agents take through your product as a subway map: one line per route, thickness is the share of sessions that took it, colour is the stage, and a route that never finished ends in an open dashed segment. The fastest correct route sits on top. Hover a line for its session count, its median, and a run you can open.
URLs: what agents read to use your product
A new org-wide URLs page, plus a URLs tab on every prompt, pulled straight out of the traces:
- Every page an agent fetched, with status, response size and duration
- The web searches they ran to find those pages
- MCP calls by provider, and CLI calls by binary
With a domain filter, sortable columns, a per-run status breakdown and CSV export. If agents keep hitting a 404 in your docs, or keep searching for a page you never wrote, it shows up here.
Traces you can actually read
- Resulting code at the bottom of every trace: the files the agent ended up with, in a file tree with syntax highlighting
- A Verified chip and goal-met marker at the top (the prompts column formerly called Reward is now Verified)
- Tool payloads pretty-printed, Codex
apply_patchrendered as per-file diffs, ToolSearch turns showing which tools loaded - Setup command output folded under the setup pre-step, and a download menu for the whole session
Also shipped in August
- Org credentials: named secrets on their own settings page, with values that never leave through any API
- An estimated wait on a queued run, from how long the last one took
- Insights deduplicate themselves now, matched on embeddings, so one friction point stops appearing five times under five names
- Edit a prompt's end goal from the configure sheet
Coming soon
Experiments: a spreadsheet for AX evals where a cell is a prompt, a run or a check, and you fill right to run it across models. Plus tags, search and bulk runs on the prompts list, and a public AX Tier List.
Thoughts?
We're always here if you have any thoughts on what we can do better. Reach out at mika@2027.dev or DM me on X anytime: @heymikasagi.