Ramp Ops

---
name: ramp-ops
description: "Use when running ops agents in the Ramp pattern: repetitive loops vs exploratory dynamic workflows, declarative correct-trace, trace-debug over benchmarks, least-privilege, and cost culture. Triggers on \"Ramp ops\", \"loops vs workflows\", \"correct-trace\", \"trace-debug\". Non-triggers: generic workflow orchestration with no ops surface (use workflows). Outcome: an ops loop system with least-privilege, cost control, and on-call-ready inspection."
metadata:
  origin: ECC
---

# Ramp Ops

Operate agents the Ramp way: cheap **loops** for repetitive toil, **dynamic workflows** for exploratory work, declarative traces for debugging, least-privilege by default, and cost culture that platformizes top spenders.

## When To Activate

- The user says Ramp pattern, loops vs workflows, babysit PR, rebase train, dead-code sweep.
- On-call, SRE, or platform work needs a digital coworker across GitHub/Linear/Slack/Datadog/Sentry/Zendesk.
- Debugging relies on slogans or benchmarks instead of real execution traces.
- Agent spend concentrates in a few workflows nobody owns.

## Workflow

### 1. Split loops vs dynamic workflows

- Loops (repetitive, stable shape): babysit PR, rebase, dead-code removal, flaky-test rerun, dependency bumps.
- Dynamic workflows (exploratory, unknown shape): incident triage, new integration, migration planning.
- Rule: if it ran the same way 3 times, convert it to a loop; if it never runs the same way twice, keep it dynamic.

### 2. Write declarative correct-trace

- Declare the expected trace first: steps, tools allowed, artifacts produced, done-criteria.
- Run the agent, then diff actual trace vs declared trace.
- A deviation is a finding, not noise: fix the declaration or fix the agent, never both silently.

### 3. Prefer trace-debug over benchmarks

- When something fails, read the full trace (inputs, tool calls, outputs, timing) before quoting any benchmark.
- Reproduce from the trace with the smallest input that still diverges.
- Benchmarks rank models; traces fix systems. Never substitute one for the other.

### 4. Enforce least-privilege

- Each loop gets minimum scopes: only the repos, channels, and APIs it needs.
- Destructive actions (merge, close, page, deploy) require an explicit gate or allow-listed target.
- Review scopes when the loop changes shape; privilege creep is the default failure.

### 5. Build cost culture

- Track cost per loop: tokens, calls, wall-clock, owner.
- Platformize top spenders first: shared cache, smaller model for the stable part, batching, fewer wakeups.
- Report weekly: top 5 loops by spend, action taken, delta next week.

### 6. Inspect with a digital coworker

- Stand up Inspect-style coverage: GitHub (PRs, CI), Linear (issues), Slack (signals), Datadog/Sentry (telemetry/errors), Zendesk (tickets).
- Add Glass-style visibility: what the agent saw, did, and skipped, per run.
- Wire on-call SRE: page on loop failure, attach the trace, keep a runbook link in the alert.

## Anti-Patterns

- Dynamic workflow for pure toil: expensive reasoning where a loop would do.
- Rigid loop for exploration: forcing incidents into a script that cannot adapt.
- Debugging by benchmark: quoting scores instead of reading the trace.
- Broad credentials for a narrow loop: one token that can do everything.
- No cost owner: top-spender loops with no name attached.
- Alert without trace: paging a human with a verdict but no evidence.

## Exemplo

```text
Toil: babysit de PRs + rebase train → loops (rodou igual 3x = converte)
Incidente novo → dynamic workflow (forma desconhecida, não força script)
Debug: lê trace real (inputs+calls+timing), não benchmark; custo/semana por loop com dono
```