Post 3

The Graveyard

Evidence. Specific failures. The bodies.

The zombie agents

Still works, kind of. Nobody knows why when it doesn't. Builder left six months ago. Slack channel dormant since March.

Produced fluent contract language that was wrong. Caught by counterparty's lawyer. 17 more found. Nobody sure they found them all.

Auto-approved everything under $5K regardless of category — including personal expenses. Nobody noticed for three months. The auditor noticed.

Standing ovation at the all-hands. Never left staging. Hardcoded API key. Builder rotated teams. Agent went silent. Nobody noticed for four weeks.

No list. Nobody can produce an inventory of every autonomous agent running. Nobody's job was the inventory.
No dashboard. Legal has no idea what Finance is running. Support has no idea Marketing built three agents off a Glean tutorial.
No evaluation. Three test runs, scans the responses, looks reasonable, smells okay, ships. A Slack message: "let me know if anything looks weird." Nobody ever let them know.
No repair path. Manual prompt edits. Hope. Ship. No score trajectory. No traceability. The loop does not close.

Evidence stage — these are the inevitable outputs of conditions from posts 1 and 2
Specific stories, not abstractions — reader recognizes their own org's versions
Named tools from Post 2 reappear (Ironclad, Coupa) — the arc connects
Doesn't diagnose yet — just shows the bodies. Diagnosis comes next (Mirror)
Credibility through honesty: "here's what actually happened"

specific stories no inventory unattended risk credibility through honesty audience: all three