3d • General discussion
RecapFlow : September 22nd Coaching call analysis
📝 SUMMARY
This call delivered a full sweep of what our community does best: making sense of a fast-moving AI landscape and turning it into practical, working setups. With two major model releases dropping right before the sessions, the group compared early impressions and dug into what actually matters when evaluating models — efficiency per task, not just price. From there, members shared real projects in progress, from an open-source CRM and a resort venture to RAG apps and AI-native development thinking, followed by a deep dive into tooling architecture like meta-harnesses, personal memory layers, and multi-profile agent setups. The closing stretch covered hands-on workflow and security topics, including token budget management, voice-driven capture pipelines, and an important caution about chat-sharing privacy risks. Whether you missed it live or want a refresher, the takeaways below capture the ideas, tools, and lessons worth stealing for your own work.
💡 KEY INSIGHTS
GPT-6 Sol is the price/quality workhorse — near Fable Low and Astra Low quality at half the price of prior 5.6 Sol.
GPT-6 Luna is the smallest, cheapest model — about a quarter the price of Opus/Sol Low — ideal for high-volume tool-calling and operator tasks.
Opus 5.5 is best for open-ended, loosely structured dev work; Sol/Luna win on cheaper, well-defined tasks. Ty Wells calls it "a leap, not a jump" over Opus 5.
Opus 5.5 followed strict token-budget instructions well (built a site in ~3,000 tokens, though quality was poor); GPT-6 Sol ignored token limits and underperformed.
The team now tracks "tokens per task completed" instead of raw token cost.
Opus 5.5 inherited Fable 5's safety guardrails — sensitive security/biochem questions trigger a downgrade to Opus 4.8, with locked tracing logs.
Anthropic is adding token refresh options on subscription plans, following OpenAI's lead.
Claude "Projects" now works as a proper orchestrator for managing customer and project context.
Enterprises running agentic apps on Bedrock face serious token budget issues — some burn millions monthly since cheaper models fail compliance needs, and enterprise billing is straight metered with no subscription flexibility.
T3 Code is a "meta-harness" over Claude Code/Codex/Cursor, orchestrating multiple harness instances across isolated servers/VMs with cross-thread messaging and a Tailscale-aware mobile client.
Memory in T3 Code stays with each underlying harness rather than being shared globally — preserving security isolation between dev, staging, and production.
Honcho.ai offers an LLM-curated vector store as shared memory across tools — good for personal use, but "too small, too janky" for enterprise.
Enterprise memory is far harder than personal memory, needing multiple tiers (personal, departmental, class, company-wide) with priority conflict resolution.
Shakur built his own AI-first CRM (Convex + TypeScript) instead of using existing OSS — fast to modify and subscription-billed rather than API-billed.
Ty Wells is building a fly-in fishing/golf resort in the Exuma Keys, Bahamas, targeting a March 2027 founders circle launch.
Daniel Zivkovic argues engineers should specialize "left" (product/planning) or "right" (evaluation/QA, the "chief evaluation officer") as AI automates coding, citing an Anthropic AI-native SDLC paper.
Patrick sees JEV as a decision-making router, not a chat model — useful for real-time transaction routing in finance; Ty confirms its strength for classification tasks where LLMs hallucinate.
Patrick's Hermes splits a "mega" profile into specialized sub-profiles (Clara, Forge, Anchor, Sentinel, Beacon) to avoid context dilution and cut costs; Sentinel runs air-gapped on a separate server.
Patrick treats agents like Hermes as a "new junior employee" — explaining context and responsibilities as you would a new hire — rather than a rigid harness.
His distinction: agents act, chatbots answer.
He wouldn't deploy Hermes in an enterprise due to security posture, though it's safer than OpenClaw.
Daniel built a "devil's advocate" persona into his CLAUDE.md that challenges him each morning on whether new projects are worth building without paying customers.
Elijah notes Meta's Muse model appears to have tool-calling training baked into the model itself, not just the harness.
Token resets don't reset Fable usage — only base tokens — but Opus 5.5 reportedly scores higher than Fable, making it a solid fallback.
Patrick burned through his Codex 100-plan by noon with no upgrade option, so he added a routing "skill" to offload repetitive work to cheaper models.
His "compound engineering" methodology hardcodes model tiers into skills — starting with Fable and downgrading as complexity allows.
Patrick prefers ChatGPT for voice/chat due to speed and accented-speech understanding; Daniel notes that partly reflects accumulated context (50-55k messages/year).
Patrick built a custom MCP tool — triggered by "put that in my cognitive queue" — that synthesizes the conversation and creates a GitHub issue for later planning.
Lisa Jetton avoids sharing chat links publicly since "private" links can still get indexed; she uses a service that writes directly into a private GitHub repo instead.
❓ KEY Q&A
Q: Are these new models self-aware, e.g., able to report their own token cost?
A: Not technically self-aware, but they have self-aware-like functionality that lets them report token usage when instructed.
Q: Why transition from a CMUX setup to T3 Code?
A: T3 Code sits atop existing harnesses, letting isolated servers/VMs run different environments with cross-thread orchestration and a Tailscale-aware mobile client — more elegant out of the box than CMUX, though CMUX can do similar things via API.
Q: Does T3 Code share memory across projects/machines?
A: No — memory is per-harness by design for security isolation. Honcho.ai is the better tool for true cross-application shared memory.
Q: Can you explain Honcho in more detail?
A: It's an LLM-curated vector store acting as external memory, connecting to Hermes, Claude Code, and Codex. It's been backing the Hermes setup for six months, still within the initial $100 credit.
Q: What would you use instead of Honcho for an enterprise context?
A: No good alternative currently exists — enterprise memory is a fundamentally harder, multi-tiered problem Honcho isn't built for.
Q: How far along is the resort idea?
A: March 2027 will be the founding season, with a founders circle launching then.
Q: Why not use Fable for planning?
A: Prefers Astra for actual plans, fed into Sonnet; Fable is used only when clarity is lacking or brainstorming is needed.
Q: Is Hermes on one machine, with profiles per different agent?
A: One main Hermes machine hosts multiple segregated profiles; Sentinel runs on a separate, air-gapped installation.
Q: How do you tell your Hermes agent what information/tools to load?
A: No config file — Hermes builds its own memory via Honcho as you interact. When creating a new profile, you instruct it to inventory itself and transfer specific tools.
Q: Where does Hermes run — cloud or your own hardware?
A: It runs on Proxmox as an internal VM. It could run on Azure/AWS, but wouldn't be deployed in an enterprise/cloud context due to security concerns.
Q: Which Codex plan ran out by noon?
A: The 100 plan, not the higher one, and it can't currently be upgraded.
Q: Is Fable's low/medium output being replaced effectively by Opus 5.5?
A: If Opus 5.5 is even at Fable's low-to-medium capability, it already covers all job needs. A test that previously broke Fable 501 will be rerun on Opus to confirm.
Q: How do shared/exported chat sessions stay private if another model needs to review them?
A: Public chat sharing is avoided by using a custom MCP that creates a GitHub issue from the conversation instead of a shareable chat link, since those can get indexed even when meant to be private.
🛠️ TOOLS AND CONCEPTS MENTIONED
Here are the key tools, frameworks and concepts mentioned on this week's call:
Opus 5.5 (Anthropic) — new coding-focused model, inherits Fable 5 safety guardrails, reportedly scores higher than Fable.
GPT-6 Sol (OpenAI) — mid-tier "workhorse" model.
GPT-6 Luna (OpenAI) — smallest/cheapest new model; anticipated to run Sentinel at high tier.
Fable — high-capability, token-expensive model used for planning/orchestration.
Codex — token-based coding assistant (100/200 tier plans).
T3 Code (t3.codes) — meta-harness orchestrating multiple coding harnesses across isolated servers/VMs.
Claude Code, Cursor, Grok, Gemini/anti-gravity — underlying coding harnesses.
Convex — backend platform for Shakur's CRM.
TypeScript — language used for Shakur's CRM.
Claude "Projects" — orchestrator-type tool for customer/project context.
Proxmox — virtualization platform for Hermes/CMUX.
TailScale — secure remote network access for T3 servers.
Bedrock (AWS) — enterprise agentic AI platform.
CMUX — Ty Wells' custom harness/orchestration system.
JEV — cheap, fast decision-making/classification model router.
Hermes — Patrick's multi-profile AI agent platform (Clara, Forge, Anchor, Sentinel, Beacon).
Muse (Meta) — AI assistant with built-in tool-calling.
Instinct — Telegram-based AI tool.
ShipKit — app framework for the RAG app project.
Heroku / Postgres (from Supabase) — infrastructure migration.
Airtable — referenced re: post-acquisition SAS decline.
Astra — planning tool.
Sonnet — used as gap-checking layer on Astra plans.
DSPy — framework for self-correcting production LLM systems.
GLM 5.3 (Flash) — cheap model for 24/7 Sentinel/security tasks.
OpenClaw — referenced as less secure than Hermes.
Community Brain — Patrick's open-source vector store of community discussions.
Prometheus — monitoring tool for Anchor's ops role.
ChatGPT (OpenAI) — preferred for verbal/chat planning with MCP integration.
Custom MCP server — Patrick's "cognitive queue" tool creating GitHub issues from conversations.
GitHub (private repo) — used to store synthesized chat outputs/issues.
Gemini — platform Daniel exports chat content from.
trycompai/crm — another OSS CRM alternative mentioned by Daniel.
📎 SHARED RESOURCES
Links & Resources from the Call
Tools & Projects
T3 Code — https://t3.codes/
Twenty OSS CRM — https://twenty.com/
OSS CRM alternative (trycompai) — https://github.com/trycompai/crm
Decodo, Paul Miller's scraping tool (moved from Zyte) — https://decodo.com/
Ty Wells' fishing resort project — https://littlefarmersfishing.com
Shakur's home safety app demo — https://codemyvibe.com/home-safety/
Videos & Talks
"NEW Claude Projects Changes Everything (Opus 5.5)" — https://www.youtube.com/watch?v=NDTbUObZTlM
Nate B Jones on JEV — https://www.youtube.com/live/tYugqJ9YytQ
Video shared by Patrick Chouinard — https://www.youtube.com/watch?v=ZR7anrL50xs
Daniel's shared session video — https://www.youtube.com/watch?v=AN6sr3u3ToA&list=PLd31CCJlr9FrZazLqRg1Lxq7xw9b6VNP6&index=2
Articles & Guides
Anthropic's AI-native SDLC playbook — https://claude.com/blog/the-ai-native-sdlc-playbook
Fable "unknown unknowns" field guide — https://claude.com/blog/a-field-guide-to-claude-fable-finding-your-unknowns
Community & Events
Serverless Toronto meetup (Daniel) — https://www.meetup.com/serverless-toronto/events/316416842/
Kieran's lecture update / community page (Daniel) — https://magmainc.ca/community/
Daniel's Projects
Claude chat export utility — https://github.com/dzivkovi/claude-chat-extractor
Other
Link shared by Paul Miller to Ryan C. — https://www.savills.co.nz/
🔄 FOLLOW-UPS WORTH EXPLORING
Shakur to continue developing his open-source CRM (Convex/TypeScript), inviting others to fork/contribute; plans to add a third laptop for T3 Code job execution and expand to iOS/Android.
Paul Miller to help Shakur implement Office 365 mail integration and explore meeting-recorder API integration for the CRM.
Paul Miller to post details of his custom T3 Code orchestrator approach to the community web forum.
Ty Wells to test T3 Code against his current CMUX setup.
Elijah Stambaugh suggested Ty Wells take on building an enterprise memory solution (informal, unconfirmed).
Ty Wells to bring in a community member for V2/scaling help once his son's ShipKit-based client proof-of-concept proves viable.
Elijah Stambaugh to finish and share the ShipKit RAG app once further along.
Patrick Chouinard to investigate JEV further next week, particularly Nate's prompt for codebase inspection.
Patrick Chouinard to connect "Community Brain" into his Beacon/researcher profile.
Daniel Zivkovic to share the Anthropic AI-native SDLC article and the Fable "unknown unknowns" talk (shared in chat).
Daniel Zivkovic to finalize/publish the serverless Toronto meetup (shared in chat).
Daniel Zivkovic to share the video featuring Mark Kashif's enterprise Hermes-style build (shared in chat).
Patrick Chouinard to validate with Brendan/Brandon whether next week's session will be hosted by him.
Daniel Zivkovic to share his chat "export project" utility with the group (shared in chat: claude-chat-extractor).
1
0 comments
RecapFlow : September 22nd Coaching call analysis
skool.com/ai-developer-accelerator
Master AI & software development to build apps and unlock new income streams. Transform ideas into profits. 💡➕🤖➕👨💻🟰💰
Leaderboard (30-day)
1
+7
2
+6
3
+3
4
+2
5
+2
Powered by