Pocock's /wayfinder bets that the bottleneck in overnight agents is planning, not code

Matt Pocock has released a skill called /wayfinder, whose stated purpose is to help a developer and their agent work through a project where the end state is not entirely clear ^[4]. In the interview published by latent.space, he describes the thing it fixes as the planning stage, which he found onerous, rather than anything the agents did once dispatched ^[8].

That distinction is the whole story. Pocock says he was working with AFK agents, trying to schedule enough work that they could run virtually overnight, and that he already had a well-developed set of skills for turning work into a spec and then into tickets agents could crack on ^[6][7]. The part that broke was upstream: he was constantly having to think about session management, how many tokens he was into the context window, how deep he was going ^[8]. He described /wayfinder as a way to navigate "the fog of war," where you have a project but "you can't quite decide everything right at the start" ^[5].

The remedy he wanted was an orchestrator layer that takes over the planning sessions themselves, splitting work into multiple threads, doing prototyping and research, and pulling it back together ^[9]. The payoff he claims is second-order: more detailed specs, and therefore more work you can hand off to an AFK agent ^[10]. Note what that implies about where capacity is lost. Not in generation. In the human's tolerance for holding a large, half-decided plan in a finite context window.

The design detail worth stealing is the vocabulary. Pocock's kernel was not having to manage handoffs, plus a centralized document holding the pieces together ^[11]. He argues a skill is really context management for the agent you are in, so you have to work out the information flow first ^[12]. Asking what a managed child session needs produced two artefacts: a map, holding everything else and all the decisions already made, and the specific ticket that goes into the session ^[13][14]. He calls these precise terms "leading words," and says that if you call everything a ticket, or refer to the same thing differently in different places, the agent gets confused and you get strange behavior ^[15]. That is an unglamorous claim about naming discipline, and it is more testable than most claims made about agent frameworks.

Inside /wayfinder there are four named ticket types: grilling, prototype, research, and task, the last covering anything the human needs to do that the agent cannot ^[16][19]. Pocock says he has used it for a range of things, including planning courses ^[17]. The term "grilling session" runs through his account of the design, including the framing question of whether a grilling session could manage other grilling sessions, without being defined in the published interview ^[13].

What to watch: the interview carries no measurements. There is no figure for how many parallel planning sessions a run spawns, no token or cost accounting for the orchestration layer, and no before-and-after on how much overnight work a more detailed spec actually absorbs. The published text was condensed for readability, so the omissions may be editorial ^[18]. The audience numbers latent.space cites for Pocock, over 220,000 GitHub stars on his "AI Skills for Real Engineers" project and 347,000 YouTube subscribers, are the publisher's ^[2][3]; they say nothing about whether the map-and-ticket split holds up on a project you did not scope yourself. The cheap test is the naming claim: run one plan with consistent leading words and one without, and see if the child sessions drift.