Skip to content

Long-term goals

Strategic direction (5–10 year horizon) for what Darcy is building toward, independent of any single project. This is the filter: when a new opportunity shows up, does it compound toward the mission and thesis below, or not? If not, default to no.

The short-range view — active projects, waypoints, 90-day focus window — lives on What-Im-Working-On and index. This page is the layer above that: why the current projects are the current projects.

Mission

Experiment with and build truth-seeking AI — mechanisms, agents, and products that help humans and take on hard problems worth solving.

That's the one-line north star. Truth-seeking is deliberate: not just fluent outputs, but AI that grounds claims, surfaces uncertainty, verifies before it acts, and stays honest about what it doesn't know. The draft → human-gate → execute → verify pattern the agent fleet runs on is one expression of that discipline. The work spans mechanisms (how models are routed, distilled, grounded, and audited — see projects/ai-model-distillation), agents (a scoped fleet — Crusty first; successors under projects/openclaw-autonomy-org — each owning a narrow slice, sharing this wiki as context), and products (apps and services where the phone or a phone line is the right trust surface). Help humans is the benefit filter: coaching, companionship, creative tools, physical-world ops — anything that makes someone's life measurably better. Hard problems worth solving is the ambition filter: not engagement hacks or wrapper theater, but interesting constraints (trust, seniors, physical nodes, durable context) where getting it right actually matters. Sustainable income, when it lands, is a consequence of solving real problems well — not the mission itself.

The mission (what we're building toward) and the thesis below (why this bet compounds) are the two filters used together. Everything active should serve both.

North-star thesis

"The winners of the next decade won't be the apps with the best UIs. They'll be the agents with the most trusted access to your data. The smartphone era rewarded app design. The agent era rewards context depth and trust." — Peter Diamandis (paraphrased from a 2026-04 newsletter)

Darcy agrees with this and is betting on it. Translated into a personal positioning statement:

Be one of the people who knows — from hands-on practice, not theory — how to build agents that (a) have deep, durable context about a specific owner's life or business, (b) are trusted with real permissions (money, accounts, hardware), and (c) stay safe enough that the owner keeps extending that trust over years.

Refinement (2026-05-01): this is not a blanket anti-app thesis. Standalone apps can still be viable income paths when the app is the permissioned client for a trusted workflow: camera, microphone, HealthKit, Photos, local storage, push notifications, playback, offline use, family/device permissions, or secure on-device processing. The weak bet is UI polish alone; the strong bet is app-as-trust-surface for a context-rich workflow or agent.

The thesis has three legs; every current active project is chosen to strengthen at least one of them.

Leg What "winning" looks like Why it's hard
Context depth The agent has a persistent, structured, up-to-date model of the owner's projects, history, constraints, preferences, and in-flight work — not re-explained each session Most people don't have a system for capturing this; LLM memory alone isn't enough
Trusted access The agent can actually do things — post, pay, reconcile, schedule, order — under scoped permissions, with an audit trail Every real integration is a trust decision with a blast radius; most people (rightly) won't grant it without proof
Durable trust The owner keeps extending permissions over months and years because the agent has never caused a quiet disaster One bad autonomous action can reset the trust curve to zero

Why this is the right bet for Darcy specifically

Not every thesis fits every person. This one does, for concrete reasons:

  • Already building the pattern for myself. projects/crusty is a working agent with scoped integrations and a Telegram human-gate; projects/personal-projects-wiki + projects/llm-maintained-context are the context-depth substrate; projects/operator-agent is the first real-money application. These exist before the thesis was named — good signal that the direction is natural, not forced.
  • Temperament fit. The job rewards people who care about low blast radius, clean interfaces between an agent and the world, and being honest about what an agent can and can't be trusted with yet. That's closer to Darcy's operating style (see personal-operating-system) than to "ship fast and apologize later."
  • Leverage over scarce skills. Most LLM-era builders know how to prompt models. Far fewer know how to build the durable context layer and the trust architecture around them. Darcy has a head start on both.
  • Physical + digital combo is rare. projects/operator-agent pushes the agent into physical trust — ops, money, hardware — (EV multifamily charging first, smart vending sprint deferred). Long ladder still runs through scaled L2 → laundromat → RaaS → own hardware — exactly where "agents with trusted access" has the most upside and the fewest competitors.

Current projects mapped to the thesis

Project Context depth Trusted access Durable trust
projects/crusty ✓ Aye Robot morning loop, session-log history ✓ Telegram, xurl, image gen, web search, scoped shell (2026-05-10: mail digest + meals off Crusty → grok.com) ✓ hardened memory, Telegram gate, session-log audit, clean-laptop minimal-permission stance
projects/personal-projects-wiki ✓✓ primary artifact — LLM-maintained structured context about every project — (read/context, not a doer) ✓ public-safe split: raw/ private, wiki/ curated-public
projects/llm-maintained-context ✓✓ the meta-skill: how to keep the context layer current without hand-editing — ✓ version-controlled, SCHEMA-governed
projects/operator-agent builds up as ops history accumulates (charger telemetry, settlement disputes, landlord threads) ✓✓ real payout surfaces + leases + tariff math under a Telegram gate — vending-era SKU PO skill parked unless automation removes restock labor the whole point — prove the gate pattern against real P&L
projects/ayerobot-comic + projects/aye-robot-crusty-paused-x-automation creative + cadence history ✓ X API v2 posting (porting), mentions polling ✓ de-risks the operator gate under boring weekly pressure first
projects/openclaw-autonomy-org umbrella framing for multi-agent context ✓ multi-instance pattern (Crusty + successors) ✓ Lobstar Wilde is the reference for what identity-stable agents look like over time
projects/ai-model-distillation — — ✓ on-device / small-model path matters when "trusted access" means "runs locally, doesn't phone home"; Boxy POC (generic video-audio thread) informed Phase 1
projects/happy-body ✓ coaching history, assessment videos, mentee progress over time ✓ native iOS video capture / upload / playback for a health-adjacent coaching surface ✓ high-trust human coaching posture; app justified only as assessment / feedback client, not generic UI
projects/lunarcast — — ✓ "CloudKit, no sign-in" is an explicit privacy / minimal-data posture — same values, different surface
projects/buddyline ✓✓ per-senior model of routines, tone, topics, engagement — compounds per-owner over months and is hard for big labs to replicate ✓✓ outbound voice calls + SMS + cron-style scheduling against a senior's real daily life under family-granted scope ✓✓ the entire product posture is a trust posture — always-identify-as-AI, no medical overreach, supplement-not-replacement, no forced app install. Hits all three legs; currently logged as a strong candidate deferred to post-2026-07-18

Everything active is at least one ✓ column. Items that would score zero across all three columns (pure UI/design plays, growth-hacking tactics, content farms) are correctly categorized as paused or out-of-scope.

Skills to deepen (the personal curriculum)

The thesis is only a bet if the skill-building plan is specific. These are the capabilities Darcy is explicitly investing in:

  1. Agent trust architecture. The draft → human-gate → execute → verify pattern, session-log audit trails, scoped credentials, blast-radius analysis, rollback paths. Practiced live in projects/crusty, projects/aye-robot-crusty-paused-x-automation, projects/operator-agent.
  2. Personal context engineering. Wiki-as-agent-memory, raw-vs-curated split, schema discipline, LLM-maintained indexing. Practiced live in this repo — see projects/personal-projects-wiki, projects/llm-maintained-context, SCHEMA.md.
  3. Agent ops / reliability. Daily loops that don't rot: memory compaction survival, key precedence discipline, session JSONL vs gateway logs (see recurring-issues), cron/env hygiene, "did the scheduled job actually run" verification.
  4. On-device / small-model deployment. Where trust means "the data never leaves the device" — WhisperKit, Core ML, distilled models. Slow-burn via projects/ai-model-distillation; Boxy (generic on-device video-audio POC) was Phase 1 hands-on practice.
  5. Physical-world agent operations. Charger/session telemetry pulls, Stripe + network settlement reconciliation, driver + landlord communications. Vending SKU PO drafts remain a paused playbook. The projects/operator-agent + projects/ev-charging-shoreline-poc line is explicitly the learning vehicle for this leg.
  6. Multi-agent composition. Separate agents with separate trust surfaces (Crusty = morning Aye Robot on OpenClaw 2026-05-10; grok.com = digest/meals; future agents for other domains), coordinated via shared documentation, not shared credentials. Framed under projects/openclaw-autonomy-org.

Positioning bets (5–10 year)

These are not commitments — they're the plausible shapes "winning" could take, organized so future-Darcy can check whether current work actually points this way. The Mission up top is explicit about fleet (plural), so the primary frame below is the fleet's roster. External outcomes and stances that aren't themselves agent seats follow underneath.

Plausible fleet roster

Each row is a candidate seat in the fleet — a scoped agent identity with a narrow domain, its own trust surface, and a defined blast radius. Seats are not promises; they're the composition that would make the mission real if every bet landed. The order is roughly "nearest to live" first.

# Seat What it owns Current prototype / prior art Trust surface (blast radius)
1 Personal ops + creative (Crusty — live) Morning Aye Robot comic loop on OpenClaw (Telegram + xurl) — 2026-05-10: only unattended daily job left on the machine. Mail digest + 3 pm meals → grok.com. De-risks the gate pattern under boring weekly pressure. projects/crusty today; projects/ayerobot-comic / projects/aye-robot-crusty-paused-x-automation. Telegram gate, X API (image-only posting), scoped shell. Low blast radius by design.
2 Physical-node ops specialist (Crusty skill surface → likely own instance) Telemetry + settlement ingestion, payout reconciliation drafts, landlord/PM correspondence, occasional driver-facing triage across the Operator Agent EV / physical portfolio (vending SKU path archived unless it returns cleanly). Income seat candidate — physical-world lane. projects/operator-agent, detail on projects/ev-charging-shoreline-poc; starts as a Crusty skill surface for Node 1, likely splits into its own OpenClaw instance once portfolio multiplies (different credentials, different trust bar from personal ops). Stripe/network settlement APIs, utility mail, EMS/charger dashboards, Telegram gate. High blast radius — strongest case for instance separation.
2b Productized-agent / AaaS surface (new 2026-04-22 — conditional on first engagement) Custom Crusty-style agents built for other small businesses — same draft → human-gate → execute → verify pattern applied to the client's surface (their CRM, their inventory, their customer-chat, their supplier inbox). Sits alongside Seat 2 as the second candidate income seat, this one in the software / consulting lane rather than physical nodes. No active engagement — lane remains; substrate = projects/crusty + this wiki + projects/llm-maintained-context. Per-client scoped credentials, repo write access, optional site-embed / widget surface. Medium blast radius — per-client isolated; never shares secrets with Seat 1 or Seat 2.
3 Wiki / context maintainer (half-built) Keeps this repo current: raw/ → wiki/ projection, SCHEMA enforcement, log + index hygiene, cross-reference maintenance. Today this is Cursor + Darcy; long-run it becomes a dedicated context agent so the fleet's shared memory doesn't rot. projects/personal-projects-wiki + projects/llm-maintained-context + SCHEMA.md. The meta-skill leg of the thesis. Repo write access only; no money, no external posting. Lowest blast radius — good candidate to autonomize early.
4 Coaching surface (conditional — personal) Async, trust-sensitive feedback — Happy Body mentoring (personal). projects/happy-body (native iOS video + web at happybodygrok.com; limited uploads for flexibility assessments). Conditional on Seats 1–3 proving the pattern. On-device / no-phone-home preferred where applicable; high trust bar. Would come after the first three seats.
5 Consumer-product voice + senior companionship (conditional) Two related shapes of the same seat: (a) the calm-consumer stance (app-store copy, support replies, restrained social cadence) for any LunarCast-style product that resurfaces; (b) the voice-first senior companionship surface proposed as projects/buddyline — phone-number-based AI companion for seniors with a family-managed dashboard. Both share the "agent, not app / meet users where they are / no aggressive discovery" posture; (b) is the first concrete candidate for what this seat could own, vs (a) which has always been a stance waiting for a product. projects/lunarcast stance as the reference; projects/buddyline (added 2026-04-23) as the first named product candidate — deferred to post-2026-07-18 revisit so it doesn't compete with Waypoints 1 + 2; projects/triviabalance shares the senior audience and is a natural content / distribution neighbour. Read-mostly consumer surface is low blast radius; voice companionship for seniors is materially higher (outbound calls, vulnerable demographic, real-time conversation), so BuddyLine specifically may justify splitting off as its own seat once prototyped — hence the seat is noted as possibly splitting if BuddyLine lands.
6 Band / gig ops (optional — may stay a Crusty skill forever) Press-kit updates, cold outreach to venues, FB + site upkeep, gig-logistics email. Currently 100% manual; lowest urgency because it's in the identity/life bucket, not the income leg. projects/rusty-cage + projects/music-production. Might never get its own seat — more likely a skill surface on Seat 1. Band inbox, FB, site. Low blast radius.

Cross-seat invariants (fleet-wide, not per-seat): shared wiki as the context layer (Seat 3's job is keeping it clean); draft → human-gate → execute → verify on every state-changing action; scoped secrets per seat (never one credential shared across seats); session-JSONL audit trail; identity stability over time (Crusty is Crusty, not a rotating model+prompt). projects/openclaw-autonomy-org is the umbrella framing for how seats compose.

Two candidate income seats, not one (2026-04-22). When the mission's practical human benefit leg needs a revenue expression, either Seat 2 (physical-node ops) or Seat 2b (productized-agent / AaaS) can serve it — plausibly both in parallel over time. Seat 2 is the physical-world lane (EV L2 Shoreline POC projects/ev-charging-shoreline-poc first → scaled ports/sites → laundromat → RaaS → own hardware; vending sprint deferred — weekly restock incompatible with automation posture); Seat 2b is the software / consulting lane (agents built for other small businesses). They're not competitors — they strengthen different sides of the thesis (Seat 2 maximizes trusted access against real P&L; Seat 2b maximizes context depth across multiple clients). Seat 2b has no active client engagement on the wiki; the lane remains for a future scoped build. 2026-05: Seat 2 diligence pivoted toward utility-grant-heavy EV installs documented on projects/ev-charging-shoreline-poc. A long-term Option 4 frame still holds when a Seat 2b engagement does land: Seat 2b funds Seat 2 — AaaS revenue buys into the physical ladder at a higher rung, shortening the path to the robotics endpoint that gives Seat 2 its actual pull.

What the roster tells me about sequencing. Seats 1 and 2/2b are where the active 90-day focus window pays off — Seat 1 has to be boringly reliable before either income seat touches a real client or real P&L. Seat 3 is a natural next expansion (low blast radius, high leverage, and it fixes the current "wiki maintenance is a Darcy bottleneck" problem). Seats 4–6 are explicitly conditional and should not compete for attention during the current focus window.

Non-seat bets (outcomes, stances, reputations)

Separate from the roster because these aren't agents — they're things that become true as a side effect of running the fleet well.

  • Pattern exporter. The personal-wiki + draft-gate-execute-verify pattern is repeatable. The wiki is already public via Quartz. Future expansion: a canonical write-up (not an info-product, a documented working pattern) that other builders fork. Already partly happening via SCHEMA.md and projects/personal-projects-wiki.
  • Trust-first product instincts. The values driving projects/lunarcast (CloudKit, no sign-in, no ads, no aggressive discovery) are the right consumer-product instincts for the agent era. Relevant even if Seat 5 never gets staffed — it's a design compass, not a roster position.
  • Lived operator credibility. A version of "operator of agent-run physical nodes" that isn't about owning a seat but about having run one — the kind of reputation you can't fake by reading about it. Earned by Seat 2 doing its job for long enough that the "I've run this" is true.

What the thesis tells me to say no to

Equally important — this is what the thesis filters out:

  • Pure UI / app-design plays whose value is "prettier than the competitor." Standalone apps remain valid only when the phone is the needed trust / permission surface; otherwise the era reward for generic app polish is ending.
  • Growth-hacking tactics (auto-like, auto-follow, follower farms). Also now mostly removed from the X self-serve API, conveniently — see projects/aye-robot-crusty-paused-x-automation and the 2026-04-17 X pricing change.
  • Dropship / content-farm / info-product models. No trust leg, no context-depth leg, no durable relationship.
  • "LLM wrapper" products with no persistent context layer and no real integrations. The moat is exactly the layers such products don't have.
  • Crypto / franchise purchases / pure consulting. Already explicitly killed on projects/operator-agent for the same underlying reason — no compounding asset.
  • Platform-dependent automation (e.g. the old consumer-web X automation path). Trust requires stable interfaces; browser automation against someone else's policy doesn't qualify.

Review cadence

  • Per-project filter. When adding or unpausing any project, check: (a) does it push the mission forward — i.e. does it advance truth-seeking AI (better mechanisms, agents, or products), genuine human benefit, or a hard problem worth solving — and (b) which of the three thesis legs (context depth / trusted access / durable trust) does it strengthen? If both answers are "none," default to no.
  • Quarterly. Re-read this page at the end of each 90-day focus window (next: 2026-07-18). Update the alignment table, prune skills that aren't actually being practiced, add new ones that are.
  • When the thesis moves. If the field shifts (e.g. agents commoditize faster than expected, or the trust leg turns out to be irrelevant), update the thesis itself here rather than quietly pretending nothing changed.

Meta

  • Created: 2026-04-20 — triggered by the Diamandis quote above.
  • 2026-04-21: Added explicit Mission ("produce an autonomous fleet of agents that provide value and generate income 24/7") at the top. The thesis was already here; this names the concrete thing being built so the per-project filter has a "what" as well as a "why."
  • 2026-04-21 (same day, v1.2): Restructured Positioning bets into a Plausible fleet roster (seats 1–6, each with domain + prototype + trust surface) plus a small Non-seat bets list for outcomes and stances. Reason: now that the mission explicitly names a "fleet," the positioning section should read as who's in the fleet rather than a flat list of strategic bets. Seat 1 = live Crusty; Seat 2 = Operator Agent ops specialist (physical-world income lane); Seat 3 = wiki/context maintainer (half-built today); Seats 4–6 are conditional.
  • 2026-04-22 (v1.3): Added Seat 2b — Productized-agent / AaaS surface as a second candidate income seat in the software / consulting lane, sitting alongside Seat 2 rather than replacing it. projects/carlot-chat is the first concrete candidate engagement and was moved from "side work, outside focus window" to "candidate top priority pending engagement confirmation" on its own page; alignment table gained a carlot-chat row. Added a "Two candidate income seats, not one" sequencing note under the roster making explicit that the two income seats strengthen different sides of the thesis (Seat 2 = trusted access against real P&L; Seat 2b = context depth across multiple clients) and that which one gets the focus-window attention depends on which produces a confirmed engagement first — Option 4 frame ("Seat 2b funds Seat 2") captured as the long-term synthesis between them.
  • 2026-04-23 (v1.4): Added projects/buddyline — voice-first AI companion for seniors — as the first concrete candidate for Seat 5. Reframed Seat 5 from "Consumer-product voice (conditional)" to "Consumer-product voice + senior companionship (conditional)" to cover both the original LunarCast-stance shape and the BuddyLine shape, and flagged that BuddyLine's materially higher trust surface (outbound calls, vulnerable demographic, real-time conversation) may justify splitting it into its own seat once prototyped. New BuddyLine row in the alignment table scores ✓✓ on all three legs (context depth + trusted access + durable trust) — the cleanest thesis fit on the roster. Explicitly not pulled into the current focus window: decision is to revisit at 2026-07-18 unless a concrete pull arrives sooner (a family member asking for it, projects/triviabalance engagement data wanting a conversational surface, or an xAI voice-pipeline shift that changes the math). Also captured a real previously-unnamed cross-project synergy: BuddyLine and projects/triviabalance share a senior audience and could share a content pipeline / per-senior profile over time. Voice-for-education parallel experiment folded into projects/education-startup-ux rather than a new project page — it's a new channel on an existing product line.
  • 2026-04-29 (v1.5): projects/carlot-chat archived — client not investing; Seat 2b lane remains, first candidate engagement closed without delivery. Alignment table and roster row updated; "Two candidate income seats" note revised so sequencing is not blocked on Carlot. Similar SMB dealer-style client work explicitly deprioritized vs other projects (What-Im-Working-On).
  • 2026-05-01 (v1.6): Refined the north-star thesis so it is not read as "apps are dead." Standalone apps remain viable when they are the permissioned client for a trusted workflow; pure UI polish stays filtered out. Added projects/happy-body to the alignment table as the clearest current example of app-as-trust-surface (native iOS video assessment / coaching loop).
  • 2026-06-04 (v1.7): projects/education-startup-ux explicitly day job, not personal portfolio — removed from personal fleet Seat 4 prototype column; Seat 4 now Happy Body (personal) only. Alignment table and on-device skill note updated.
  • 2026-06-06 (v1.8): Mission reframed from autonomous fleet / income 24/7 to truth-seeking AI that benefits humans on hard problems. Fleet, gate pattern, and income seats remain as implementation layers; per-project filter and income-seat note updated to match.
  • 2026-06-06 (v1.9): Removed projects/carlot-chat and projects/education-startup-ux from the wiki entirely. Boxy retained only as a generic on-device video-audio experiment on projects/ai-model-distillation (no use-case detail). Alignment table, Seat 2b/4, and income-seat note updated.
  • Status: v1.9. Treat as a strategic compass, not a plan. The plan is What-Im-Working-On.