The agent-tooling layer is maturing fast, and this cluster is the practical end of it: multi-agent coding without worktree chaos, hosting, and skills you can drop into Claude Code today.
David Wells’ Multi-Agent Coding Without Worktree Chaos is the most useful workflow piece I’ve read in a while. He went all-in on git worktrees for multi-agent coding, then drifted back to a single branch, because worktrees defer collisions to merge time rather than removing them. His answer is coordination, not git ops: plan the work, interrogate the model for non-obvious questions, turn the plan into a directed task graph, then let a swarm pick up unblocked nodes while a coordinator (“ring leader”) loops every few minutes. Collisions are prevented by explicit file reservations and an agent mailbox for blockers, not by git status archaeology. The line that stuck: “opening more terminals is not a system.” This is structured parallelism, and it’s clearly where the leverage is.
Which makes Herdr the natural companion: “tmux for coding agents.” It’s a Rust binary (no Electron, no account, no telemetry) that runs your agents in real PTYs on any box, keeps them alive when you close the laptop, shows blocked/working/done/idle state at a glance, and lets you reattach over SSH from your phone. Real panes, a plugin marketplace, and a CLI + socket API so agents can orchestrate the multiplexer themselves. If you’re running the kind of swarm Wells describes, this is where it lives.
Devashish’s Two Qwen3 Models on One DGX Spark is the deep-systems entry for anyone chasing local inference. The goal: run Qwen3-Next-80B-FP8 for heavy work and Qwen3-4B for fast turns coresident on one Spark (119.67 GiB unified memory), behind a LiteLLM proxy on a single endpoint. It took a weekend, and the write-up is the “residency math” pain: gpu_memory_utilization is a snapshot against total card memory, not free memory; Qwen3-Next is mostly Mamba, so KV-cache demand is dominated by Mamba state alignment, not attention KV; and coresident processes race: they don’t negotiate. Niche, but if you’re eyeing a Spark to own your inference, this is the hard-won map.
Two Agent Skills close out the cluster, both of which land right on your workflow. Josh Pigford’s SEO Content Generation skill is a self-contained content engine: drop it in ~/.claude/skills/, and each run researches what to publish next, picks the content type, writes one deeply-researched on-brand piece, and registers it in a dedup ledger, built for 2026 search, optimized to rank in Google and get cited by ChatGPT/Perplexity. Given your own recommendations series, the “decide → research → write → verify → register” loop is worth studying even if you never run it. And unslop-text is the one I think you’ll actually want: a skill that strips the cues that make prose read as AI-generated (the em dash, “it’s not just X, it’s Y,” sycophantic openers, delve/tapestry diction, the “in conclusion” wrap-up), and warns against the newer over-corrected “trying not to sound like AI” register (forced fragments, fake typos, bolted-on swearing), which it treats as its own tell. It’s grounded in a Reddit analysis of what readers actually name as giveaways, ships a deterministic scanner for CI, and is honest that the real tells (uniform rhythm, empty fluency) are structural and need a human ear. This is essentially your HARDLOCK ruleset as a portable skill: required reading for how you edit The First Legion.