The full session summary

Agentic App Architecture

• Core concept: agent lives behind the app; the app is a surface for the agent to project onto
• Surfaces: web app, mobile, WhatsApp, chat, browser extension
• Agent paints the surface, updates it on a schedule or on demand
• Analogy: like the early app-store moment, but for agentic software
• Agent adapts output to user expertise level
• Expert users: dense signal
• Novice users: agent detects and adds explainer layer
• Fire-tracking agent as open-source example
• Polls sources every 2 hours or on manual refresh
• Renders a widget: mile-radius circles, fire locations, wind direction/strength
• Available on GitHub; suggested as a starting point for the host’s dashboard

Pete’s Dev Workflow and Toolstack

• Full stack on Cloudflare: instantly scalable, startup credits ($10k available via their program), best pricing
• Flu Framework: Pete’s open-source agent framework built on Cloudflare abstractions
• Astro team’s new framework also maps cleanly to Cloudflare; Pete recently switched to it
• Cloud Code as primary harness; orchestrates Opus 4 + GPT-5.5 in parallel
• Opus handles strategic/design thinking; GPT-5.5 handles strict coding tasks
• Work Trunk abstraction over git work trees; no manual git operations
• Two new plugins changing the game:
• Impeccable: design system generator; breaks AI-slop aesthetic patterns, produces Apple-tier output
• Ponytail: “senior engineer with a ponytail”; picks right libraries, prevents over-building, cuts token use
• Ship plugin: end-to-end flow from design to PR merge to deploy
• Presents output as HTML to Pete as “product manager,” not raw code
• Uses image gen for docs; opens and merges PRs automatically
• Auto-dream / retro loop: at end of each session, agent reviews process improvements
• Sends a PR back to Pete for approval if a workflow change is found
• Status bar in Cloud Code shows current pipeline stage; hand icon signals when input needed

Agent vs. Harness vs. Loop: Conceptual Clarity

• LLM = raw token stream; cannot call tools directly
• Harness (Cloud Code, Codex, OpenClaw, Pico): intercepts special tokens, fires tool calls
• Black-box harnesses (Cloud Code, Codex) hide their system prompts; can conflict with user intent
• Agent = harness + body: file system, compute, sandbox, internet connection
• OpenClaw on a Mac Mini = agent; OpenClaw alone = harness
• Two loop types Pete recently distinguished:
• Cadence loop: agent runs continuously on a metric, improving incrementally (Kaizen-style)
• Build loop: give agent an 80% target per pass; iterate to completion without micromanaging
• Agents can share memory without querying each other’s LLM
• One agent reads another’s HTML wiki / file system directly (“telepathy”)
• Faster and cheaper than routing through a second LLM call
• Auto-dream (Anthropic, leaked): nightly session-log review, memory tree indexing
• Pete’s Home Zero agents dream into HTML wikis navigable by both agents and humans

Model Landscape and Cost Trajectory

• GPT-5.5: strong coder, terse, less conversational than Opus; good for pure coding tasks
• Opus 4: preferred for strategic thinking, conversation, orchestration
• GPT-5.6 (upcoming): three-size lineup (Soul, Tara, Luna)
• Soul: Fable-equivalent, priced at current Opus 4 cost
• Tara: Opus 4 equivalent at ~half the current Opus 4 price
• Token efficiency: solves problems at roughly one-third the tokens of current models
• GLM 5.2: near Opus 4 on benchmarks at 6-10x lower cost; not yet a daily-driver replacement for coding
• Fusion on Open Router: reportedly achieving Fable-level results; Pete hasn’t tested yet
• Thesis: coding capability is approaching a ceiling; intelligence gains will differentiate models going forward
• In 6 months, cheaper models will handle current coding tasks; expensive models will pull ahead on other dimensions

Nick’s AI Trading Firm

• Goal: AI-native firm trading real capital on Coinbase autonomously
• Architecture centers on “The Cleaner”: approved systems agent that executes live trades
• Catches all signals 24/7; solves the human sleep/distraction problem
• Supporting agents:
• Edge Reaper: scrapes forums for trading systems, back-tests against historical data
• Signal Forge Factory: parses systems into JSON, runs analysis pipeline
• Risk Officer: governs live-money gates before The Cleaner acts
• AI Cost Governor: allocates daily token budget per agent; emergency kill switch on command center
• Command center GUI: toggle switches, agent status, report generation
• Self-improving agent concept: trades 3x/day live, enters research mode overnight, reviews its own trades, updates approach
• Current workflow: ChatGPT Pro as project manager, Codex as software engineer
• Every Codex prompt generated by ChatGPT Pro; Nick never writes Codex prompts directly
• New thread started daily due to context drift; documents evolve with each thread reload
• Next phase: once v1 package is complete, drop entire codebase into GPT-5.6 for rapid v2 refactor

Host’s Trading Dashboard Project

• Observational agent: watches price and order book in real time + historical analysis
• Double-zero theory: price held at even dollar (e.g. $59,820.00) signals market-maker continuation of downward momentum
• Agent to help develop and validate further methodologies
• Educational layer: secondary goal, much further out
• Target customer: already has financial literacy (candles, averages, basic instruments)
• Recommended onboarding path: babypips.com preschool through middle school only
• Avoids teaching indicator-stacking that doesn’t hold up over time
• Key differentiator: order book data brokers don’t collect or surface historically
• No broker with hundreds of thousands of accounts provides this; intentional omission
• Human-in-the-loop philosophy: agent augments trader judgment, does not replace it
• Infrastructure: currently on AWS; hit billing overages; thresholds and alerts now in place
• Pete suggested migrating to Cloudflare to avoid surprise costs

Home Zero: Real Estate AI Platform

• Repositioned away from “next Notion for real estate” toward AI-native home search
• Three-agent architecture:
• Home Advisor: user-facing agent for home shoppers
• Market Agent: tracks listings, news, market trends across a region
• Home Agent: property-level intelligence
• Target market: “Claude Dads” (and moms): technical users already running Cloud Code
• BYOK model: users paste their own API keys (GPT-5.5 or similar) for their advisor agent
• Pete operates market and home agents; users pay their own inference costs
• Pricing: $15/month tier; $30/month tier; inference cost target under $15/user/month
• Structure: public benefits company; mission to drive housing costs toward zero
• Go-to-market: YouTube videos with Tony Laverde (local GC/musician) on AI for homeowners
• Foreclosure strategy marketplace: modular agents for niche strategies, upcharged ~50% on inference
• Planned Cloud Code integration: users can operate Home Zero directly from their Cloud Code session

Paonia Town Politics: Stefan Wynn and Board Accountability

• Stefan Wynn (former town administrator) filed police reports in March characterizing citizens’ speech as threats
• Reports included Pete passing out flyers, Kaja posting on a community Facebook thread
• Police found no basis; Wynn submitted resignation letter the same day, citing fear for his family
• Resignation letter sent from his town email account, as town administrator, not as a private citizen
• Records obtained via CORA request; key findings:
• Wynn forwarded emails to his attorney, then to police; redactions were improperly applied (PDFs with removable black boxes)
• Attorney correspondence suggests he sought a cease-and-desist against Pete’s flyer distribution
• Wynn named Kaja as “accomplice” to flyer distribution; she did not distribute flyers
• Board response: no public acknowledgment; no agenda item; no response to Pete’s formal complaints
• Kala Rose tabled a generic free-speech affirmation rather than addressing the specific conduct
• Sheree told Pete she may not be able to discuss it
• Pete’s position: this is a board oversight failure, not just a Stefan Wynn issue
• Standard set: if board doesn’t address it, next administrator can do the same without consequence
• Samira (town clerk) added two backdated personnel incidents to Kaja’s record the same day Kaja went to the mayor
• Pete’s ask: one special meeting, open dialogue (not a 3-minute comment window), on the record

Town Governance Reform

• Town administrator role: only 25% of salary funded by the town; 75% from water, trash, sewer utilities
• Argument for splitting the role: operational utility management vs. board-staff liaison are different competencies
• Treasurer role should also be separated (new administrator Sund agreed at last meeting; audit firm expected to require it)
• Town attorney: paid ~$1,000/meeting to attend; host counted 27+ “ums” in one meeting; considered low-value
• Pete’s suggestion: train a non-attorney town staffer on procedures at ~$40/hour instead
• HR manual: $10,000 contract awarded to Indiana firm; bid at $7,500 to stay under discretionary threshold
• Pete already published an AI-generated manual using comparable Colorado towns as source material; available for free
• Suggestion: CORA-request the Indiana draft, diff it against Pete’s version with AI
• Parking minimums: Pete drafted a reform plan eliminating minimums; Mancos (same size) already adopted similar policy
• Broader pattern: host notes a decade-long cycle of administrator issues, including a prior $400k embezzlement
• Framing for council: evaluate the role structure while the position is vacant; best time to make changes
• Pete’s resolution criteria: one special meeting, curiosity from board, named individuals (Kaja, Zane) cleared on record

Next Steps

• Set up Slack "agent maker" channel and invite the group (Peter)

Avoids context-switching and lets Pete share workflow updates without repeating conversations one-on-one.

• Share Flu Framework link and session recording with host (Peter)

Host wants access to the Cloudflare agent framework and the transcript from this session.

• Schedule technical interview prep session with host

D's panel interview (infrastructure, security, whiteboard) is one week from tomorrow; host is free anytime including later today or Thursday.

---