AI platform work dies when “everyone owns it” or when a single embedded researcher owns eval for the company. Topology matters: platform team for gateway/eval rails; embedded partners for product UX.
⚡ TL;DR: Platform owns gateway, budgets, tracing, prompt registry, shared eval harness. Product squads own feature prompts, golden sets, and UX. Hire for security + systems + ML eval literacy — not demo-prompting only.
Ownership matrix
| Capability | Platform | Squad |
|---|---|---|
| AI Gateway | ✅ | consume |
| Cost dims / budgets | ✅ | set limits |
| Feature prompt | templates | ✅ |
| Golden eval set | harness | ✅ cases |
| Tool IAM | ✅ patterns | request |
Platform (2–6 eng) ── rails ──► many squads
Embedded AI eng (optional) sit in high-stakes squads
Hiring signals
- Shipped production systems with SLOs
- Can read IAM and threat models
- Comfortable with eval statistics
- Skeptical of vanity accept rates (Day 42/56 vibe)
❌ Hiring only “prompt engineers” with no systems background to own the gateway.
Failure modes
Platform team becomes a ticket sink for every prompt tweak. Publish self-serve prompt registry docs. Embedded hires without production ownership create demo culture — pair them with a systems mentor.
Closing checklist
- [ ] RACI published
- [ ] Eval ownership explicit per feature
- [ ] Platform staffing matched to gateway load
- [ ] Career path for AI platform eng
- [ ] Quarterly topology review
Series navigation
Day 95: Incident Response When the Agent Goes Wrong · Day 97: Build vs Buy the Agent Runtime
Last updated September 11, 2026
Discover more from CheatCoders
Subscribe to get the latest posts sent to your email.
