Very important agents (Friends)

Summary of Very important agents (Friends)

by Changelog Media

1h 38m•December 5, 2025
Listen to Episode

Overview of Very important agents (Friends)

This episode of Changelog and Friends is a freewheeling conversation about agents, AI tooling, and the developer ecosystem. Hosts and guest Nick (the “Very Important Man”) trade news, product recommendations, tooling tips, and opinions — from Tiger Data’s Agentic Postgres to Anthropic’s acquisition of Bun — and dig into how agents are changing how we write and ship code.

Topics discussed

  • Agentic Postgres (Tiger Data) — a Postgres variant built for agents (native retrieval, zero‑copy forks, MCP server + CLI, free tier).
  • Anthropic acquires Bun — what it means for Bun’s future and the JS ecosystem.
  • The rise of agents as developer assistants and the shift from manual coding to AI‑assisted production.
  • Claude, Claude Code, CloudCode, Opus 4.5, Gemini 3, ChatGPT and other model comparisons/experiences.
  • Plugins/skills/marketplaces for agent ecosystems (Nick’s personal marketplace and examples).
  • Notion Agent, Raycast AI, and practical voice/shortcut automations (Shortcuts + ChatGPT).
  • Browser/“agentic browsing” debate — will the web be replaced by chat/agent UIs?
  • GitHub concerns (performance, sponsorships, actions slowness) and projects moving to alternatives (Zig → Codeberg).
  • Sourcegraph / AMP split into two companies.
  • Personal tech: Vision Pro, Sora (AI video app), OmniFocus, Home Assistant, Linux vs macOS development workflows.
  • Ads/partners mentioned: Fly.io, Tiger Data, Namespace, Notion, NordLayer.

Key takeaways

  • Agents are changing developer workflows: they “call, retrieve, parallelize, and plug into infrastructure” — reducing repetitive coding work and increasing output, but requiring new guardrails and review habits.
  • Agentic Postgres aims to be a database built specifically for agents (large vector/text workloads, sandboxes, migrations) — worth exploring if you’re building agent workflows.
  • Anthropic’s acquisition of Bun keeps Bun open source (MIT) and with the same team, promising better integration with AI coding tools and faster releases; community reaction is mixed.
  • Use-Case best practices: have agents do boilerplate/minutiae and keep human developers focused on business logic, architectural decisions, and code review.
  • Tooling matters: faster CI (Namespace), integrated agent UIs (CloudCode), and workspace automation (Notion Agent, Raycast) noticeably improve developer productivity.
  • The web and search are likely to change as LLMs surface answers; site owners should anticipate new discovery models (AI recommendations vs classic SEO).

Notable insights & quotes

  • “Agents are the new developers.” — framing the episode’s thesis: agents increasingly replace repetitive human work in coding workflows.
  • Nick’s practical advice: set agent/system prompts/output styles so the model handles routine code generation while you retain control of business logic and design.
  • On Bun acquisition: Bun remains MIT-licensed and public; Anthropic promises to speed releases and optimize Bun for AI tooling.
  • Practical workflows are evolving: “You’ll produce more code but write less of it yourself.”

Practical recommendations & action items

  • Try Agentic Postgres if you’re building agent-driven apps that need heavy text/vector search and quick sandboxing: see TigerData.com.
  • Speed up slow GitHub Actions with Namespace (drop-in, caches dependencies, Docker layers, build artifacts).
  • If you use Claude (or Claude Code / CloudCode), experiment with:
    • Output styles (system prompt templates) to steer agents toward teaching, reviewing, or automating mundane tasks.
    • “Handoff” patterns: let subagents open new context windows when approaching context limits to avoid losing thread context.
  • Build small agent skills (examples Nick mentioned):
    • Code simplifier: auto-refactor boilerplate without breaking behavior.
    • Multi-model consultant agent: query multiple models (Grok, Codex, Perplexity, Gemini, Claude), collate answers, deliberate, and produce a consensus.
  • Use Raycast AI / launcher for quick AI lookups (developer productivity).
  • For voice and home automation: explore Home Assistant and Shortcuts + LLM integrations as the place where agentic voice assistants may first become practical.

Quick how-to (high level)

  • Agentic Postgres: visit TigerData.com → install CLI → spin up service (three-command flow, per Tiger Data promo).
  • Make an agent useful in your workflow:
    1. Define output style (system prompt) that specifies when the agent should be conservative and when it should “auto-complete” boilerplate.
    2. Create agent/skill that encapsulates repeated tasks (e.g., code cleanups, tests, or multi-model research).
    3. Use subagents for parallel research and a final “deliberation” pass to synthesize answers.
  • Speed CI: replace or augment slow GitHub Actions steps with Namespace for caching and faster runs (one-line change in many workflows).

Companies / products mentioned (with quick notes)

  • Tiger Data — Agentic Postgres (DB for agents: native search, zero-copy forks, MCP server).
  • Anthropic — Claude, CloudCode, buyer of Bun.
  • Bun — JS runtime, now acquired by Anthropic (remains MIT).
  • GitHub & Codeberg — migration discussions (Zig project moved to Codeberg).
  • Namespace — faster CI, caches dependencies & artifacts.
  • Notion Agent — AI teammate integrated into Notion workspaces.
  • NordLayer — enterprise VPN/zero-trust sponsor mention.
  • CloudCode / Claude Code — Anthropic tools for coding agents.
  • Sourcegraph / AMP — recently split into separate companies.
  • Raycast — quick AI launcher / “Google now” replacement for devs.
  • Sora — AI video composition/novelty app (kids can be upset by realistic clips).
  • Home Assistant — open source home automation (possible voice/agent frontier).
  • OmniFocus / Things / Todoist — personal task managers discussed.
  • Zed, NeoVim, Deno, Go, Convex, TanStack — various dev tools/editors/runtimes mentioned.

Predictions and closing notes

  • Prediction: developers will produce more code but manually write less of it; human role becomes more about oversight, design, and product judgment.
  • Practical tip: use model/system-level settings (output styles) to customize agent behavior to your workflow — let the AI handle the boring bits and keep humans on the strategic parts.
  • Episode is conversational and exploratory — useful for developers and managers wondering how to integrate agents, what products to watch, and how to preserve code quality and ownership as AI tooling accelerates.

If you want the main items to investigate first: Agentic Postgres (Tiger Data), Anthropic + Bun developments, Namespace for CI speedups, and experimenting with output styles/handoff patterns in your agent stack (CloudCode / Claude or equivalent).