CSM Health & Renewals Agent

The time saved.

The AESOP build produced a deployable full-stack scaffold — the agent, the build approach wired end to end — in a single pass. Here’s what that pass was worth, measured against a competent engineer doing the same from a clean PRD by hand.

12–19 hrs
of senior-engineer time, compressed to minutes. One to two working days of scaffolding, gone.
01 · The math
ArtifactManual hoursWhat it is
Risk engine + escalation logic4–6Six flag types, compound-risk detection, escalation rules
Supabase schema2–3Seven tables, RLS on every table, per-role policies
Next.js dashboard4–6Typed fetch layer, severity-colored flags, escalation badge
Deploy config + infra docs1–2Dockerfile, Fly config, pinned requirements, README
Setup prompts + diagram1–2Repo setup, Supabase, architecture — runnable
Total12–19One run of the build, not one sprint
02 · What it saved
The boilerplate

Everything a competent engineer writes before the interesting part. Routing, schemas, auth plumbing, deploy wiring, the dashboard shell. All of it is decided-and-done, not decided-and-then-built.

The decisions

The platform choices are made for you. FastAPI on Fly, Next.js on Vercel, Supabase for data and auth, a provider-flexible LLM. No bike-shedding the stack — the scaffold already picked it.

03 · What it did not save

The number is worthless if it’s inflated. Three places this scaffold is not finished work — they’re explicitly stubbed, and they’re the parts that make the agent real:

1

Conversation extraction

conversation_signals → hardcoded empty

No Gong or Salesforce fetch. The weekly “what did the customer say” is stubbed to nothing — the actual data plumbing is still to build.

2

LLM narration

summary · next_action → template strings

The provider path exists, but the narrative is template text, not model calls. The “language is the model” half isn’t wired yet.

3

Eval metrics

grounding_pass_rate → hardcoded 1.0

Grounding is faked to perfect, drift detection is a stub. Real evaluation — the honest-eval principle — is downstream of the scaffold.

04 · The honest read
AESOP eliminated the boilerplate, not the product

The 12–19 hours is scaffold time — the part that was going to be slow no matter what. The remaining work — live data sources, real grounding, tests — is what the estimate deliberately does not claim to save. That’s the distinction worth keeping straight: the build didn’t skip the hard parts, it just stopped making you pay for the easy ones.

05 · Read on