Wayfinder: decision orchestration before implementation
Bottom line: Sebastian should see this. Wayfinder is not another coding agent; it is a decision-resolution and context-management layer for large, ambiguous work. It turns one rough brief into a dependency-aware graph of research, grilling, and prototype tickets, then hands a decision-complete specification to isolated implementation agents. The design closely matches moyn's emerging multi-agent workflow, but the demo also exposes substantial operational friction around worktrees, local environments, permissions, completion detection, and context cost.
Source: Matt Pocock, “LIVE: The /wayfinder Demo”, 1:14:40, accessed 2026-07-26.
Transcript: YouTube English auto-captions retrieved through yt-dlp-compatible residential proxy; minor caption errors preserved below.
What Wayfinder is
Wayfinder evolves a single long “grill me” interview into an orchestrator for work that spans multiple planning and coding sessions. It accepts a rough conversational requirement, inspects the codebase and prior art, defines the destination and scope boundary, identifies the unresolved “frontier of decisions,” and creates a linked map of tickets. Each ticket is routed to the cheapest suitable mode: a human grilling session, autonomous research, a runnable prototype, or a blocked follow-up. [Source: video 00:28–02:35, 13:13–18:12]
The intended delivery chain is:
Wayfinder → to spec → to tickets → implement → code review → human review
Wayfinder tickets resolve uncertainty; implementation tickets later reify those decisions in code. Pocock explicitly separates disposable prototype work from production delivery: prototypes inform the specification, but the specification is what gets shipped. [Source: video 11:19–11:39, 1:00:51–1:01:03, 1:12:24–1:13:55]
How the demo works
Pocock starts with a deliberately loose request: extend his Course Video Manager with a low-friction short-form video creator and investigate publishing to TikTok and YouTube. Wayfinder explores the repository, asks about the data model and UX, incorporates an older Remotion implementation as prior art, and creates a parent GitHub issue plus nine decision subissues with blocking relationships. [Source: video 00:50–15:41]
Several sessions then run in parallel. Research agents investigate publishing APIs while interactive sessions resolve product and architecture choices. When words are insufficient—for navigation, recording, and editing interactions—Wayfinder raises the fidelity by generating runnable UI variants. Human feedback selects and iterates on those variants, and resolved findings are written back to tickets with links to deeper artifacts. [Source: video 17:01–23:11, 26:41–42:19]
The map can evolve during exploration: one decision ticket may spawn a new prototype or research ticket, close itself with findings, and unblock the next decision. The map offers orientation, tickets retain concise findings, and branches/worktrees hold detailed artifacts—progressive disclosure rather than one enormous prompt. [Source: video 32:39–37:23]
The demo ends with eight of nine decision tasks complete. Pocock shows a prior Wayfinder-derived specification that became six isolated implementation tasks and an approximately 5,000-line pull request, with one agent session per ticket and aggregate review afterward. [Source: video 1:06:49–1:10:41]
What is genuinely strong
- Decision work and implementation work are separated. Unresolved product choices do not silently leak into coding tickets. [Source: video 1:13:32–1:13:55]
- Dependencies are explicit. Research, decisions, prototypes, blockers, and downstream handoffs form a graph rather than a flat checklist. [Source: video 15:10–18:12, 54:00–55:02]
- Fidelity is adaptive. Conversation handles simple questions; runnable prototypes answer “how should this look or behave?” [Source: video 27:29–28:32, 36:25–42:19]
- Preferences and values become first-class inputs. Pocock preserves hard-won Remotion caption styling and argues that agents misalign when they fail to ask what users value and which trade-offs matter. [Source: video 53:27–53:49]
- Research can discover blockers before code is written. TikTok's audit process and mandated UX cause a cheaper architectural pivot toward Buffer/Zapier instead of a premature direct integration. [Source: video 1:03:57–1:06:03]
- Context stays traceable. Downstream agents can traverse map → ticket → research/prototype rather than receiving one flattened mega-spec. [Source: video 1:08:07–1:10:11]
- The tracker is replaceable. GitHub Issues is only the demonstrated backend; Pocock says Linear, Todoist, Jira, or another system can serve as the system of record. [Source: video 1:10:55–1:11:18]
Caveats visible in the demo
- Human leadership is mandatory: early agent recommendations and one proposed taxonomy were wrong and required correction. Pocock's framing is explicit: the human is the lead; the agents are juniors. [Source: video 04:56–05:19, 12:55–13:10, 21:09–24:48]
- The first prototypes ignored real reducers and shared UI components until redirected toward the existing editor. [Source: video 30:55–33:48]
- Local execution was brittle: environment variables, installs, symlinks, routes, ports, permissions, and worktree state all caused friction. Pocock calls worktree handling “brutal” and prefers remote sandboxes or VMs. [Source: video 29:17–31:13, 38:55–40:46, 49:41–49:59]
- The chosen prototype remained incomplete: posting did nothing, video clipping remained, browser-level testing was largely absent, and the posting interaction was unresolved. [Source: video 56:35–58:36]
- Prototyping is context-heavy; agents repeatedly reread and modify many files, creating very large contexts. [Source: video 50:43–52:52]
- Parallel grilling sessions can overlap and require manual dependency linking and explicit handoffs. [Source: video 54:00–55:02]
- Completion detection is weak: users may need to inspect subissues or ask whether Wayfinder is done. [Source: video 1:14:05–1:14:24]
- The demo stops before this feature reaches implementation, so it demonstrates decision convergence—not end-to-end delivery quality. [Source: video 1:06:49–1:07:08]
Recommended viewing points
- 00:28–04:04 — concept, destination, and scope boundary
- 11:19–18:12 — full workflow and ticket map
- 26:41–42:19 — prototypes as higher-fidelity decision tools
- 50:43–55:51 — values, context cost, dependencies, architecture decisions
- 1:00:51–1:06:03 — prototype/spec boundary and external blocker discovery
- 1:08:07–1:13:55 — traceability, prior implementation evidence, and complete handoff sequence