Owns the outcome. Never touches the deliverable.
subagentroles.com catalogs the coworker tier -- WHAT job a coworker does. This site catalogs the tier immediately above it: the manager, which owns an OUTCOME (a goal plus acceptance criteria), decomposes it into typed tasks, dispatches them DOWN to exactly one coworker, and reports UP to the operator with a single summary message. It never executes the deliverable itself -- every concrete instance below has an explicit stop-and-dispatch rule for the moment it catches itself doing so.
pattern is the abstract manager-tier shape, grounded in what both concrete instances implement in common. instance rows are real, installed manager skills read live this session -- project-manager and design-manager -- each citing the exact SKILL.md it came from.
Neither manager accepts a completion on green tests alone. Both require: the coworker/subagent ran Sentry-instrumented, the ci-cd rubric cleared, and the outcome's evaluator.pass_if holds -- see each row's durability_gate field.
Both instances escalate fail/block/gate-miss the same way: a DurableTask back to the originating queue, plus an escalate envelope delivered to the operator's phone via iMessage. Neither uses Slack or Telegram -- stated explicitly in project-manager's own SKILL.md.
The two concrete instances, live
All 3 rows are live in this Worker's own D1 database (manager_tier_facts), seeded from two real installed skills read this session: anthropic-skills:project-manager and anthropic-skills:design-manager. Browse the board → or hit the JSON API.
What is real vs. what is not
Real: all 3 rows, each citing the exact skill file it was grounded in -- no invented manager instances, no fabricated queue names. This repo currently has exactly two manager-tier skills installed (project-manager, design-manager); no other domain (engineering, data, finance, legal, HR, sales, operations) has a manager-tier skill today, per this session's own skill list -- see the concept page for the honest scope note.