Retreat to attack (Friends)

Summary of Retreat to attack (Friends)

by Changelog Media

1h 44m•November 14, 2025
Listen to Episode

Overview of Retreat to attack (Friends)

This episode of Changelog & Friends (hosted by Adam and Jared) is an extended, grab‑bag edition focused on how AI/agents are reshaping development, tools, and infrastructure — plus culture notes about unplugging. Highlights include agent‑native Postgres, mesh networking for off‑grid communication, compression/archiving choices (zstd + tar), new developer roles (FDEs), agent workflows and best practices, and a debate about whether frameworks (notably React) are becoming the “substrate” for LLM-driven app creation.

Key topics covered

  • Agentic Postgres (Tiger Data): a database built with agents in mind — native retrieval/search, instant zero‑copy forks, multi‑cluster server controls and CLI tooling.
  • Alive Internet Theory (Spencer Chang): a counter to the “dead internet” idea — a nostalgic human-centric site that surfaced by decade‑based Internet artifacts.
  • MeshTastic + LoRa: off‑grid, low‑power mesh networking using LoRa radios for long‑range, decentralized comms.
  • Depot.dev (guest Kyle Galbraith): pipeline and CI/CD challenges in an era where agents massively boost code velocity; importance of faster, more reliable builds and observability.
  • Compression & archiving: issues with 7z on Linux prompted a switch to zstd + tar (“zarch”) — big speed and CPU wins.
  • Agentic/code workflows: practical tips for using AI coding agents (rebooting context, maintaining a context file, agent governance).
  • Augment Code: recommended daily‑driver tool for AI-assisted development (CLI + IDE integrations).
  • Forward Deployed Engineers (FDEs): rising role focused on customer‑facing AI deployment, integration, and success — communication + translation skills emphasized.
  • “Dead framework” theory vs. React as substrate: discussion on how LLMs and tools are hard‑coding React as the default target, making it effectively the platform many tools optimize for.
  • Personal/cultural notes: “retreat to attack” concept (unplugging to recharge), phone/watch disconnection, and sobriety/analog trends.

Main takeaways

  • The shape of developer work is changing: agents dramatically increase code output, shifting bottlenecks from writing code to integrating, reviewing, building, and operating it.
  • Infrastructure and tooling must evolve: databases, CI/CD, and build systems need to be agent‑aware (faster builds, better forks/sandboxes, native search for retrieval).
  • Practical agent hygiene matters: keep context windows fresh (reboot agents when they hallucinate), maintain explicit context files/guides, and validate outputs — humans remain essential as reviewers and product visionaries.
  • Choose efficient formats/tools: zstd (Zstandard) is a modern compression algorithm offering strong compression ratios and faster throughput than many older options (significant real‑world speed-ups reported).
  • New roles will rise: technical roles that combine engineering chops with communication, customer empathy, and AI‑translation skills (FDEs) are increasingly in demand.
  • Framework dynamics: React’s dominance is being reinforced by tooling and agent defaults; this may freeze certain patterns even if alternatives exist — there are pros/cons to that consolidation.

Notable quotes & stats

  • “Agents are the new developers. They don't click. They don't scroll. They call. They retrieve. They parallelize.”
  • On human persistence: “The internet will always be filled with real people looking for each other, answering calls for help and sharing laughs…”
  • Rough metrics cited: ~50% of new articles/blog posts in a recent study are generated by AI; MeshTastic has a reported long‑range record of 331 km (hardware/antenna dependent).
  • Zstandard sample numbers from discussion: compressed ratio examples showed zstd outcompeting some alternatives with much faster compress/decompress behavior in practice (real archives went from ~20+ minutes down to a few minutes).

Action items / recommendations

  • Try agentic infra where appropriate:
    • Explore Tiger Data’s Agentic Postgres for agent‑heavy, vector/text retrieval workloads (TigerData.com).
    • Evaluate Depot.dev if CI/CD/build latency is a bottleneck (depot.dev).
  • If you archive large datasets or want faster, less CPU‑heavy archiving, assess Zstandard + tar (“zarch”) for your pipelines.
  • Experiment with off‑grid comms for remote sites: research MeshTastic and LoRa hardware if you need decentralized, long‑range messaging.
  • Improve agent workflows:
    • Maintain a context.md and a small context‑guide to steer agents.
    • Reboot or clear agent context when hallucinations compound.
    • Use tools like Augment Code (augmentcode.com) to centralize agent interactions across CLI/IDE/Slack.
  • Learn customer‑facing AI skills: communicative problem solving, integration, and product translation — consider the FDE skillset if you want to be at the intersection of AI product and adoption.
  • Write an agent: the hosts encourage listeners to build a simple agent to learn the tradeoffs and empathy required.

Guests, projects & links mentioned

  • Guest: Kyle Galbraith — co‑founder & CEO of Depot.dev (discussion on builds, CI, agent‑era bottlenecks).
  • Upcoming guest: Spencer Chang — creator of the Alive Internet Theory site (scheduled follow‑up episode).
  • Tools & projects:
    • Tiger Data — Agentic Postgres (agent‑first DB)
    • Depot.dev — faster builds and pipeline tooling
    • MeshTastic — mesh networking software on LoRa radios
    • Zstandard (zstd) — modern compression algorithm (used instead of 7z)
    • Augment Code — AI developer platform/agent tool (Auggie CLI)
    • Fly.io — episode sponsor / hosting partner
    • NordLayer — sponsor (promo: nordlayer.com/thechangelog, code changelog-28 for 28% off yearly plan before 2025‑12‑10; 14‑day money‑back)
    • Redis — mentioned in context of FDE use cases (vector search)

Practical notes & short checklist for devs

  • If agent outputs seem off: clear agent context → provide curated context.md → re‑instantiate agent.
  • For build pipelines anticipating high agent activity: prioritize caching, parallelization, and faster builders (avoid 20+ minute builds).
  • For archiving big projects: benchmark zstd+tar vs your current approach; zstd commonly reduces CPU time and run duration.
  • Want hands‑on? Pick one small project and:
    • build a tiny agent to automate a repetitive task;
    • replace a dependency (e.g., left‑pad‑level utility) by prompting an LLM to write a small, focused implementation; or
    • spin up a simple ZSTD archive workflow for backups.

Episode wrap & community

  • Hosts cover what they’d do next: write an agent, test zarch/archiving flows, explore MeshTastic for off‑grid communication.
  • Community: changelog.com/community (Zulip) — listeners invited to discuss episode topics and follow‑ups.
  • Upcoming schedule teased: Spencer Chang interview (Wednesday) and Chris Benson (drones & robotic swarms) on Friday.

If you want this summarized into a one‑page checklist or an actionable plan (e.g., steps to migrate archives to zstd, or an agent onboarding checklist for teams), say which one and I’ll produce it.