↓Skip to main content
Zeitgeist — a spike by Chris Gathercole
  1. Creators/

Dan Disler — IndyDevDan — Agentic Engineer

About #

A software engineer with over a decade of industry experience who has focused since early in the generative AI wave on practical, battle-tested patterns for AI-assisted and agentic coding. His YouTube channel and “Principled/Tactical Agentic Coding” courses cover multi-agent orchestration, spec-driven development, and autonomous (“AFK”) coding workflows built around tools like Claude Code.

2026-06-29 — GLM-5.2 vs MiniMax-M3: Opus Has REAL COMPETITION (Model Stacking) #

YouTube · YouTube

  • Benchmarks open-weight models GLM-5.2 and MiniMax-M3 against Opus 4.8 and lightweight Qwen3.6-35B
  • Finding: each pricing tier down is roughly 5x cheaper while only marginally less capable
  • Introduces a three-tier model stack — state-of-the-art, workhorse, lightweight
  • Open weights provide substitutability/resilience if proprietary models get deprecated; true local ownership still needs ~$50-100k in GPU hardware
  • Takeaway: “Don’t pick a model — pick a MODEL STACK”

2026-07-06 — SEE CMUX SOLVE Multi-Agent Orchestration (Claude Code and Pi Agent) #

YouTube · YouTube

  • Learns Cmux live, a terminal multiplexer giving agents programmatic control over panes and sessions
  • Orchestrates 8 parallel agents across 4 teams with full visibility and flat (non-hierarchical) communication channels
  • Demonstrates three-tier orchestration: orchestrators prompt leads, leads prompt specialists
  • Core principle: “An agent you can’t SEE is an agent you can’t improve”
  • Takeaway: agentic access (API/programmatic control) is a prerequisite for a tool being usable at agent scale

2026-07-13 — FORGET Loop Engineering. Agentic Engineering is about THIS #

YouTube · YouTube

  • Reframes agentic engineering beyond “loop engineering,” centering on three actors — engineers, agents, and code — working inside software factories
  • Scales from a simple engineer-agent-review pattern up to full multi-agent orchestration with scout, plan, build, and test agents in isolated sandboxes
  • Key claim: engineers still sit at the two fixed constraints — planning at the start, reviewing at the end — while added compute buys confidence in between
  • Goal: build systems that build systems, templating expertise into reproducible workflows
  • Takeaway: invest effort in the planning/review bookends, not in tuning the loop itself

2026-07-20 — Engineers… STOP Picking GPT-5.6 Sol OR Claude Fable 5… FUSE THEM #

YouTube · YouTube

  • Argues for model fusion — running multiple state-of-the-art models in parallel instead of picking a single one
  • Builds a fusion harness pairing Claude Sonnet 5 and GPT 5.6 Tera, then escalates to Claude Fable 5 and GPT 5.6 Sol
  • Workflow: slash commands gather multiple model perspectives, consolidate results, then run auto-validation loops where one agent authors proofs of completion and the other builds against them
  • Takeaway: “Two state-of-the-art models beats one. Two models that FUSE the best of both worlds beats two”

2026-07-27 — Is Anthropic STEALING Your Data? (While You PAY FOR IT) #

YouTube · YouTube

  • Investigates whether Anthropic uses customer data for model training, contrasting consumer vs. commercial tiers
  • Argues users effectively “pay twice” — once in cash, again by handing over IP that shapes aggregate, anonymized product-development signals
  • Claims incentive alignment, not goodwill, is what actually protects privacy in a vendor relationship
  • Introduces an “AI Sovereignty Ladder” framework for protecting proprietary agent work from data-retention risk
  • Takeaway: check the data-retention policy for your specific tier before treating a subscription as a safe harbor for IP