Sherwood Callaway is the founder of Sazabi, an AI-native observability platform built for engineering teams that ship fast. Before this, he started the observability team at Brex in 2019 and spent his career in DevOps, infrastructure, and site reliability, which means he has a deep love-hate relationship with tools like Datadog. In this episode he lays out a three-part manifesto for tearing down traditional observability and rebuilding it around agents.

We get into why dashboards and the three pillars are on the way out, how Sazabi uses git-backed agent memory and per-thread sandboxes, the MCP versus CLI debate, GEPA-style parallel agents, and the brutal reality of building evals for an agent that lives in a sandbox.

What we cover:

- Logs are all you need: Why Sazabi kills metrics and traces and reconstructs them from logs on the back end, so instrumentation is just a print statement.

- Chat is the best UX for observability: No flame graphs or service maps, just ask an agent and get answers in Slack.

- Monitors are dead: Replacing static thresholds with AI-generated, non-deterministic alerts plus recommended remediation.

- Read-only by design: Why the agent draws a hard line at code generation but can still kick off Cursor or Claude Code fixes and open Linear tickets.

- MCP vs CLI: Why CLI tools baked into the model weights avoid the context rot that MCP servers create, and the skills-bloat trap.

- Git-backed agent memory: Markdown memory in a self-hosted git server, a branch per thread, and merge conflicts resolved by a background workflow.

- Sandboxes per thread: Cloning your repo into the sandbox to correlate a log line back to the exact commit and author.

- GEPA and parallel agents: Spinning up many agents on the same input, the missing reduce step, and natural selection for prompts.

- The Big Eval psyop: How far you get with context engineering and a good harness before you actually need evals.

- Eval pain in a sandbox: Fuzzy-matching bash commands, spinning up AWS and Postgres dependencies, and a TypeScript monorepo with Turborepo.

- Security stances: Read-only, no public internet access, whitelisted domains, and guarding against prompt injection.

If you build AI agents, run production systems, or are rethinking your observability stack, this one is for you.

Links and Resources:

- Sazabi: https://sazabi.com

- Sherwood Callaway on LinkedIn: https://www.linkedin.com/in/sherwoodcallaway/

- SiliconANGLE on Sazabi: https://siliconangle.com/2026/04/08/startup-sazabi-bets-logs-ai-agents-replace-traditional-observability-stacks/

- GEPA (Genetic-Pareto prompt optimization): https://github.com/gepa-ai/gepa

- Model Context Protocol (MCP): https://modelcontextprotocol.io

Timestamps (approximate, adjust before publishing):

00:00 The mission: dethrone Datadog

01:30 AI for observability, not observability for AI

03:15 The three-part manifesto begins

03:25 Thesis 1: stop looking at your telemetry, chat instead

04:45 Thesis 2: logs are all you need

06:20 Vibe observe and simpler instrumentation

06:55 Thesis 3: monitors are dead

08:25 AI-generated alerts and remediation

09:30 Read-only by design: drawing the line at code gen

10:50 MCP vs CLI and the context rot problem

13:55 Skills bloat and progressive disclosure

15:10 Sandboxes, the Sazabi agent, and the log DB tool

19:25 Git-backed memory and shared state across threads

24:30 GEPA, parallel agents, and natural selection for prompts

30:50 Evals and the Big Eval psyop

34:45 Why evaling bash commands is so hard

41:40 Security: read-only, no internet, prompt injection

44:00 Advice for AI engineers, hiring, and the beta

#Observability #AIAgents #DevTools

Logs Are All You Need: Rethinking Observability with AI Agents

MLOps.community

Logs Are All You Need: Rethinking Observability with AI Agents