Zero-Touch Weeks, Secret AI Coaching, and Self-Healing Security

He ran his HVAC company for a week without touching a single tool

Oscar Hidalgo

Twelve agents, one orchestrator, zero manual tool use for a full work week

The problemEvery tool in Oscar's business still needed him personally to open it, log in, and do the work, even though he already had a working crew of AI agents.

Oscar challenged himself to do zero manual work for a full week, running his HVAC company, his AI-consulting clients, and his own course prep all at once, using a twelve-agent crew under one orchestrator on Hermes. He has been auditing his entire stack for tools that work agent-first: he switched his email provider to Superhuman because its connector reaches every agent he runs, tested a new browser built so agents can open their own tabs without taking over his computer, and moved every agent onto a remote server, synced to his second brain with Syncthing, so they keep working even when his own laptop is off. For a client whose single agent was burning tokens re-pulling the same e-commerce data every week through slow browser automation, Oscar recommended a cheaper pattern: pipe every data source into ClickHouse once through a connector like Composio or Airbyte, so the agent only has to build the dashboard front end afterward instead of re-fetching the same data on every run. On the HVAC side, Housecall Pro runs his service technicians and Connecteam organizes thousands of daily jobsite photos automatically into per-building folders, while QuickBooks feeds a full six-month P&L forecast that gets checked against submitted technician hours every night.

You are my operations agent, modeled on the zero-touch week shared at the Agentic Society Boardroom (August 2026). I want to run my business for a week without personally opening a single tool. Interview me for: (1) every recurring task I currently do by hand and roughly how often, (2) which tools I use for each one and whether they have their own connector or MCP, (3) any task where I am re-pulling the same data repeatedly. Then: (1) propose one narrowly-scoped agent per recurring task, naming each by the role it plays, (2) flag any tool that has no agent connector and suggest an agent-first alternative, (3) for repeated data pulls, propose piping the source into a database once instead of re-fetching it live each time. Never touch a connector or credential without my explicit approval first.

Takeaway A generalist agent guesses which tool to reach for. A narrowly-scoped agent with one connector never has to guess, which is what actually saves tokens and time.

He secretly automated his call-center coaching, and his team never noticed

Chance Forman

The problemChance's KPIs, project tracking, and content production were scattered across separate tools, and his customer-service team had no consistent daily coaching.

Chance built one dashboard that runs his whole day across a portfolio of companies. A Today page surfaces anything blocked and waiting on a human's approval, and a new project cannot move past the idea stage until a short template, definition of done, goals, risks, success metrics, is filled out, which he often completes conversationally while driving. Every project card carries its own built-in chat assistant already loaded with that project's full context, running on Codex when he types and Claude when he talks. Team members get their own connected agents inside the same dashboard, and he plugged Composio into the back end so a new hire can link their own accounts with one click, no API keys required. On the content side, raw daily footage batch-uploads into the dashboard, which drafts a script and several hook variations, builds a shot list, and hands the recording to either AI or his editor for a final polish pass, tracking which version performs best on each platform. Every night, anything logged in the dashboard sweeps into a second-brain archive, so every tool he opens the next morning already has the full context of the day before. The detail that got the biggest reaction in the room: his customer-service reps believed Chance was personally listening to and grading their calls every single morning. In reality, GoHighLevel pulls each rep's call recordings automatically, an agent drafts feedback on their best and worst calls, and it goes out under his own name, until he told them the truth.

You are my operations dashboard architect, modeled on the system shared at the Agentic Society Boardroom (August 2026). Help me design one central place to run my business. Interview me for: (1) the KPIs I actually check day to day, (2) how new ideas or projects currently get greenlit, (3) whether my team already records customer calls. Then propose: (1) a short intake template that has to be filled out before any idea becomes an active project, (2) a plan for giving each project its own AI assistant with that project's context preloaded, (3) if calls are recorded, a way to have an agent draft coaching feedback on the best and worst calls automatically. Always tell me before anything goes out under my name without my review.

Takeaway The best AI rollout for a non-technical team is often the one they never notice, because it's built into a system they already use, not bolted on as a new tool to learn.

Security-fix SLA: three days to minutes

The problemRecurring cybersecurity audits generated a constant stream of findings that used to take expensive outside consultants days to triage and fix.

Jay built a suite of agents that ingests every audit finding, triages it, proposes a fix, and automatically opens a ticket in GitHub, a build Roman had challenged him to take on a few weeks earlier. A human review gate approves the fix before the agents implement it and deploy it to a staging environment, operating under a least-privilege model that only requests elevated access when a specific change genuinely needs it. Task routing runs through a model router that picks between Claude, Codex, and other models depending on the job, defaulting to a mid-tier model for most fixes and a cheaper model for heavier data-processing work, and rarely reaching for the most expensive tier at all.

You are my security-remediation agent, modeled on the pipeline shared at the Agentic Society Boardroom (August 2026). I run recurring compliance or security audits that generate a stream of findings. Interview me for: (1) where my findings currently get tracked, (2) what access levels exist in my infrastructure today, (3) my tolerance for automated fixes versus ones that need review. Then: (1) triage each new finding and propose a specific fix, (2) open a tracked ticket automatically, (3) wait for my explicit approval before implementing anything, (4) once approved, implement the fix under the minimum access level the change requires, and request more only if the approved change genuinely needs it. Never expand your own access without asking first.

Takeaway A self-healing security pipeline still needs exactly one human checkpoint, before a fix ships, never before it's proposed.