galsapir

Guardrails V2 — What We Built

Branch: DS-1233-agent-flow-management
Date: 2026-03-29 → 2026-03-31
Commits: 224f62d → c8abfa2 (5 commits)

---

New Guardrail Definitions

galsapir

Guardrails V2 — Implementation Review & Proof of Work

2026-03-29T14:17:37Z by Showboat 0.6.1

What Was Built

Branch: DS-1233-agent-flow-management
Commits: 344094f (initial) + review fixup (pending)
Date: 2026-03-29

galsapir

Adding Guardrails, Evals & Agent Flows — Operational Guide

2026-03-29T12:09:08Z by Showboat 0.6.1

Context

Branch: DS-1233-agent-flow-management
Date: 2026-03-29
Authors of the system: Itamar Apel (primary), Gal Sapir (reviewer/deployer)

galsapir

ed peers
A "C" means: around the 40-60th percentile (average)
An "F" means: worse than ~80% of peers (bottom 20%)

This approach:
- Accounts for age-appropriate expectations
- Avoids arbitrary cutoffs
- Provides meaningful context

### Component Score Calculation

galsapir

Terminal Setup: Ghostty + tmux + micro

A modern terminal workflow for engineers who SSH into remote machines daily.

What We Set Up

1. Ghostty (Terminal Emulator)

Config gist

galsapir

# Font
font-family = "JetBrainsMono Nerd Font"
font-size = 14
font-thicken = true

# Theme
theme = Catppuccin Mocha
background-opacity = 0.95
background-blur-radius = 20

galsapir

SkyPilot EC2 Cheatsheet

Connect

From terminal (after aws sso login):

ssh ec2_skypilot

From VS Code:

galsapir

GitHub Auth Setup — SkyPilot Shared EC2

This machine uses a shared ec2-user account. To avoid stepping on each other, each user gets their own SSH key + worktree-scoped git config.

Prerequisites

• SSH access to the machine (via AWS SSM)
• A GitHub account with access to the PhenoAI org

galsapir

# SDK Health Agent — Data Flow Trace

## Architecture Overview

```
User (browser) → WebSocket → FastAPI → SDK Agent → Claude CLI subprocess → Bedrock
                                ↕                          ↕
                            Langfuse                  user_data/ori/*.md
```

galsapir

Session Summary — 2026-02-11

The Big Insight

The Claude Agent SDK is an entire agent runtime, not an API call.

It's not chat.completions(). It's not a single LLM call where you send a prompt and get a response. The SDK spawns a subprocess — a full Claude Code CLI process — that has its own tool loop, its own file system access, its own multi-turn reasoning. You send it a message and it goes off and does things: reads files, globs directories, thinks, retries, reads more files, and eventually comes back with an answer.

This is the point that really crystallized today. You can read the docs and think you understand it, but until you watch the agent autonomously chain 3 tool calls to find and analyze health reports — in a subprocess you don't control — it doesn't fully land. The implications: