Map: AI adoption ladder — climb 2→3 (plugins fleet)

[kyle-sexton]

Destination

A PLAN.md-ready contract set for climbing the fleet from step 2 to step 3 of Boris Cherny's adoption ladder (artifact, Jul 16 2026), plus a charter decision on the autonomous-runner repo. Graduates to: /planning:interview (Brief per work package) → /planning:architect. Build work graduates to /work-items, never onto this map.

Notes

• Discovery + evidence: .work/ai-adoption-ladder/ — gap-map.md (synthesis), explore-plugins-inventory.md, RESEARCH-product-surface.md, RESEARCH-sandcastle-pocock.md, RESEARCH-headless-agents.md, RESEARCH-peer-frameworks.md, interview-checklist.md (Q1–Q4 resolved)
• Fleet position: solid step 2; 1→2 essentially complete; gating weakness = autonomy of the loop
• 1→2 residue (e2e in CI, pre-approved allowlist, review auto-invoke, classifier tuning) is build work → file via /work-items, not map items
• wayfind-semantics question (Pocock census flag) is plugin hygiene, off this map → /work-items triage

Decisions-so-far

Pointer index only — each decision's durable home holds the text.

• Q1 universal framing → .work/ai-adoption-ladder/interview-checklist.md (Q1 entry)
• Q2 artifact home → interview-checklist.md (Q2 entry)
• Q3 gastown/sandcastle concepts-only → interview-checklist.md (Q3 entry)
• Q4 plugins-fleet-first scope → interview-checklist.md (Q4 entry)
• No step-skipping constraint → gap-map.md § Position
• Wayfind label taxonomy grammar → github-iac#150 / fix(planning): align wayfind label taxonomy with colon-space axis grammar #235
• Trigger layer = adapters → work-item queue contract → Where does the autonomous-kickoff trigger layer live? #240 resolution comment
• Security review = default-on tool-agnostic policy contract (2 layers, per-layer blocking knob) + guided setup + audit/adjustment layer → Does security review become a default blocking gate, and in what form? #241 resolution comment
• Sandbox bar = isolation ladder L0–L3, uniform L2 floor for unattended, fail-closed → Runtime sandbox bar for step-3 autonomy #245 resolution comment + design-threads.md T2
• Guardrail matrix = 5 risk classes × 5 guardrail columns, C2-only earned auto-merge, escalation directional pending Autonomous-runner repo charter: timing and scope boundary #244 → Per-work-class guardrail matrix #243 resolution comment + design-threads.md T3
• Runner charter = charter now / build on trigger; queue-contract boundary (runner drains, never decides); self-run primary substrate, hosted via adapter capped human-gated → Autonomous-runner repo charter: timing and scope boundary #244 resolution comment + design-threads.md T4
• Return accounting = in scope, lightweight convention: human-attested counterfactual + manual-effort band at autonomous-class task close, machine cost/lifecycle layer joins; agent never estimates return fields (revisit trigger: models proven capable); aggregation deferred to Unify session and CI verification telemetry? #247 → Return accounting: in scope, and in what shape? #246 resolution comment + design-threads.md T5
• Telemetry unification = yes, contract layer: standards-only OTLP pinned to OTel CICD/VCS semconv (RC) + one custom work-item attribute (the Return accounting: in scope, and in what shape? #246 join key) + TRACEPARENT propagation across trigger→CI→agent; sink deployment-owned (existing stack | file-artifact query-on-read free default | backend opt-in); local pipeline demoted to one conforming instance, native-vs-hook duplication audit files post-graduation → Unify session and CI verification telemetry? #247 resolution comment + design-threads.md T6
• Routines = scheduled trigger adapters behind the queue contract (never a private merge path); catalog = 39-class/5-group taxonomy with judgment/output/access axes mapping to matrix, merge policy, isolation floor; v1 subset = repo-scoped proven classes, rest deferred with access triggers (incl. computer-use actuation); hosting = deployment-owned binding (invariants + decision inputs only, no default); ubiquitous language locked (loop/schedule/routine/goal) → Standing scheduled routines: which maintenance classes, hosted where? #242 resolution comment + design-threads.md T7

Not-yet-specified

• Runner escalation / ask-a-human UX (the seam sandcastle lacks) — seam fixed by Autonomous-runner repo charter: timing and scope boundary #244 charter (severity/class-routed per matrix); UX sharpens at design stage
• Return-accounting capture surface details (prompt wording, band values, tracker field vs comment shape) — convention fixed by Return accounting: in scope, and in what shape? #246, join key fixed by Unify session and CI verification telemetry? #247 (work-item attribute); details sharpen at design/architect stage
• Telemetry contract details (custom namespace naming, semconv version pin, CI emission mechanics, sink adapter shapes, native-vs-hook audit scope) — contract fixed by Unify session and CI verification telemetry? #247; sharpen at design/architect stage
• Org-enablement track shape — trigger: day-job adoption becomes concrete
• Cross-team ownership at review speed (CODEOWNERS, SLAs) — trigger: multi-team adoption
• Wayfind label-name config seam — trigger: adopting org with a non-colon-space grammar
• Security-review delivery packaging (wiring-vs-advisor split per capability, plugin category/topology, composition model) — sharpens at design/architect stage per Does security review become a default blocking gate, and in what form? #241

Out-of-scope

• 3→4 items until the 2→3 trust loop is earned (fuzzing, prod-feedback remediation, cross-repo migration engine)
• Org-enablement collateral (deferred track above)
• Vendor-hosted closed-loop (hosted agents keep a human-merge-gate by design; step 4 is self-operated)

---

Reference: Boris Cherny, "Steps of AI Adoption" (Google Doc)

[kyle-sexton]

Destination coherent — all 8 decision items resolved and closed (#240 trigger layer, #241 security review, #242 routines, #243 guardrail matrix, #244 runner charter, #245 sandbox bar, #246 return accounting, #247 telemetry unification). Contract set recorded in the Decisions-so-far index (durable homes: each item's resolution comment + design-threads.md T1–T7). Handed to /planning:interview — Briefs per work package (packaging/topology and the deferred design-stage details land there) — then /planning:architect. Remaining Not-yet-specified fog lines carry into their tagged design/architect stages. Backlog seeds filed post-graduation via /work-items: telemetry native-vs-hook duplication audit (per #247), 1→2 residue items, wayfind-semantics reclaim, fleet dogfood build work.

[kyle-sexton]

Source anchors (added for durability — the design threads re-anchor to this artifact):

• Boris Cherny, "Step & your role" (Jul 16, 2026) — Google Doc (reliable copy): https://docs.google.com/document/d/1R91ayvj7uvlxgNi--__2-Bf3w8x5r1nF-xIBN7ds8Ns/edit
• Original claude.ai artifact (intermittently erroring as of 2026-07-17): https://claude.ai/code/artifact/bfdfaef9-bc62-4dfe-ba9e-c58a26c9accf