#agentic-ai

Steve Yegge introduces Gas Town, an opinionated orchestrator for running 20-30 Claude Code instances at once, built on his Beads issue tracker. Coding agents are organized into seven roles (Mayor, Polecats, Refinery, Witness, Deacon, Dogs, Crew) that pick up persistent "molecule" workflows from a Git-backed queue, self-restart across crashes and context limits, and merge work through a dedicated queue — explicitly compared to Kubernetes and Temporal, but optimizing for task completion rather than uptime.

Argues that "orchestrator" workflows in agentic coding create a paradox of supervision — effectively supervising an agent requires the very coding skills that atrophy from over-relying on it — and adds vendor lock-in and unpredictable token costs as compounding risks. Proposes demoting AI to a secondary role: use it for planning and research, but keep manually coding 20-100% of implementation and never generate more than can be reviewed in one sitting.

I love the line that a higher level of ambiguity is not a higher level of abstraction, and the Jeremy Howard quote at the end. It ties back to the "dark flow" article I have saved — Rachel Thomas's gambling-addiction parallel to vibe coding, the same slot-machine-lever feeling.

Hadley Wickham surveys why tightly-scoped tools don't scale for coding agents, and walks through three layers of protection — explicit permission prompts, OS-level sandboxing, and a secondary LLM safety review — explaining the tradeoffs and failure modes of each.

The current state of the art described here — sandbox by default, then LLM review for anything the sandbox rejects — is what Posit Assistant just implemented. The permission-prompt approach is dismissed as security theatre: approval fatigue means users click yes without scrutiny. The narrowly-scoped safe tool approach fails for a different reason: agents perform better with fewer, more general tools than a large menu of limited ones.

Computerphile's Mike walks through why agentic coding sessions burn so many tokens: the entire context window re-enters the model on every forward pass, and each file read by a coding agent adds thousands of tokens that compound across every subsequent step. A two-file bug-fix demo accumulates ~55–60k tokens; a six-prompt starfield screensaver hit 2 million input tokens.

The GitHub Copilot flat-fee-to-per-token switch is the clearest case study here — the old model was just a subsidy. The tire-wear analogy for measuring productivity in tokens is apt. Most agentic use cases are still hard to justify on pure cost grounds outside of very targeted, short-context tasks.