I Built 8 AI Agents in 11 Minutes to Run My Business (Claude Managed Agents)

Source: Luke Carter | https://www.youtube.com/watch?v=ho2oqagDkGg Duration: 12 min | Published: 2026-04-09 Processed: 2026-04-10

Core Concepts

Anthropic has just released "Managed Agents" — a Claude console feature that lets you describe, create, and run persistent named agents directly in your own product via API, without having to build the agent infrastructure (memory, tool registry, session management, runtime) yourself.
- Before this release, builders were hacking persistence together using things like "Open Claude" (a community project that let you harness a Claude subscription programmatically). Anthropic killed Open Claude four days before shipping Managed Agents, making Managed Agents the official replacement.
- Notion is one of the reference customers — Notion's Agents feature is built on Claude Managed Agents. The pitch is "why build on Notion when you can build the same capability into your own system".
The power of Managed Agents comes from the combination of three things plugged together:
- Managed Agents runtime (Anthropic hosts and runs the agent loop).
- Custom tools exposed to the agent (leads, emails, funnels, articles, CRM endpoints — whatever the builder wires up).
- A brand wiki as persistent context (voice, community, monetisation, audience, design, frameworks, identity).
The brand wiki concept is explicitly credited to Andrej Karpathy's "second brain" idea — a living, linted, structured knowledge base that every agent reads before acting.
- Lint = automated consistency/quality checks over the wiki, run weekly, to keep the knowledge base clean and usable by the agents.
- Every spun-up agent inherits the wiki as context, so every output is automatically on-brand and business-aware.
The demonstration is an in-house CRM (the "Digital Home back end" from part 2 of the series) with a "Growth Panel" command centre UI. Inside it:
- A "Spark" agent receives natural-language requests like "check all funnels on allaboutbrands.com and tell me opportunities", then explores the CRM, reads offers, reads subscribers, reads published articles, and returns a planned funnel — including deciding which interactive funnel type best fits.
- The agent can then actually build the funnel inside the CRM (it has a funnel-builder tool exposed), pulling testimonials from the wiki and offer data from the CRM automatically.
The full agent roster shown in the demo includes: article writer, growth advisor, client onboarding, funnel architect, lead manager, email campaign drafter, wiki linter, command-centre orchestrator. Eight agents spun up in a single session.
Each agent has:
- A system prompt auto-generated from the brand wiki and CRM context at creation time (visible and editable in the Claude console).
- Tool access scoped to what it needs (funnel API, email API, lead API, wiki read, etc.).
- Persistent memory across sessions (the agent remembers prior interactions).
Autopilot vs human-in-the-loop: email sequences can be fully automated (agent drafts and sends) or held for manual review/edit — the CRM exposes both modes per workflow.
The full ecosystem being pitched: content calendar + article writing agent + growth panel + lead management + email sequencing + knowledge wiki + agent roster, all inside one owned codebase connected by a shared Supabase database and hosted on Cloudflare.
The philosophical frame: "digital sovereignty" — own the infrastructure, own the data, own the agents, don't rent them from GoHighLevel / Notion / a SaaS CRM. Managed Agents is positioned as the piece that finally makes this practical for non-enterprise builders.

Buildable Ideas

Brand-wiki-backed agent factory — a reusable scaffolding where any new client gets (a) a brand wiki populated from intake, (b) a set of Claude Managed Agents auto-wired to it, (c) a CRM shell exposing the standard tools. Onboarding compresses to under an hour.
Command-centre UI pattern — a single chat panel inside the operator dashboard that speaks to every agent. Replaces jumping between tools. Fits Himiko's desired UX.
Wiki lint agent — a weekly scheduled agent that reads the wiki, checks for stale entries, inconsistencies, missing links, and either fixes or flags them. Directly applicable to Jo's Karpathy-style memory wiki (already seeded).
Growth advisor agent — reads CRM, offers, and wiki; proposes funnels and campaigns; can execute them. Could be ported into Polerie as a packaged consulting tool.
Funnel architect tool — the tool the agent calls to actually build a funnel in the CRM. Generalisable: any deterministic builder (landing page, email sequence, lead magnet) becomes a tool that any agent can invoke.
Agent per CRM verb — client onboarding, lead qualification, email drafting, article writing, pipeline scoring. Each is one agent with one system prompt and one tool cluster. Cheap to add, composable.
Managed Agents migration path for Himiko — Jo's Hermes/Jin/Himiko stack currently rolls its own agent runtime. Claude Managed Agents is now the sanctioned backend and removes that maintenance burden.

Key Takeaways

Claude Managed Agents is now the default execution layer for any Claude-powered agent in your own product — it removes the need to build agent infrastructure (memory, loop, tool registry) from scratch.
The brand wiki is the critical IP. Without persistent, structured context every agent is generic. With it, every agent is effectively a specialist hire. Jo's memory wiki work is directly on the right path.
A single CRM with a shared Supabase DB + a roster of Managed Agents + a brand wiki replaces GoHighLevel, Notion, ConvertKit, and the typical SaaS stack for a services business.
The UX pattern that makes it usable is the command-centre chat panel — one conversational interface that routes to the correct agent and exposes results in the dashboard.
This video is the direct next step for Atlas — the architecture from part 2 (front end + back end + Supabase + GitHub Actions) is now extended with a managed-agent layer, and the stack is production-ready.

# I Built 8 AI Agents in 11 Minutes to Run My Business (Claude Managed Agents) **Source:** Luke Carter | https://www.youtube.com/watch?v=ho2oqagDkGg **Duration:** 12 min | **Published:** 2026-04-09 **Processed:** 2026-04-10 --- ## Core Concepts - Anthropic has just released "Managed Agents" — a Claude console feature that lets you describe, create, and run persistent named agents directly in your own product via API, without having to build the agent infrastructure (memory, tool registry, session management, runtime) yourself. - Before this release, builders were hacking persistence together using things like "Open Claude" (a community project that let you harness a Claude subscription programmatically). Anthropic killed Open Claude four days before shipping Managed Agents, making Managed Agents the official replacement. - Notion is one of the reference customers — Notion's Agents feature is built on Claude Managed Agents. The pitch is "why build on Notion when you can build the same capability into your own system". - The power of Managed Agents comes from the combination of three things plugged together: - Managed Agents runtime (Anthropic hosts and runs the agent loop). - Custom tools exposed to the agent (leads, emails, funnels, articles, CRM endpoints — whatever the builder wires up). - A brand wiki as persistent context (voice, community, monetisation, audience, design, frameworks, identity). - The brand wiki concept is explicitly credited to Andrej Karpathy's "second brain" idea — a living, linted, structured knowledge base that every agent reads before acting. - Lint = automated consistency/quality checks over the wiki, run weekly, to keep the knowledge base clean and usable by the agents. - Every spun-up agent inherits the wiki as context, so every output is automatically on-brand and business-aware. - The demonstration is an in-house CRM (the "Digital Home back end" from part 2 of the series) with a "Growth Panel" command centre UI. Inside it: - A "Spark" agent receives natural-language requests like "check all funnels on allaboutbrands.com and tell me opportunities", then explores the CRM, reads offers, reads subscribers, reads published articles, and returns a planned funnel — including deciding which interactive funnel type best fits. - The agent can then actually build the funnel inside the CRM (it has a funnel-builder tool exposed), pulling testimonials from the wiki and offer data from the CRM automatically. - The full agent roster shown in the demo includes: article writer, growth advisor, client onboarding, funnel architect, lead manager, email campaign drafter, wiki linter, command-centre orchestrator. Eight agents spun up in a single session. - Each agent has: - A system prompt auto-generated from the brand wiki and CRM context at creation time (visible and editable in the Claude console). - Tool access scoped to what it needs (funnel API, email API, lead API, wiki read, etc.). - Persistent memory across sessions (the agent remembers prior interactions). - Autopilot vs human-in-the-loop: email sequences can be fully automated (agent drafts and sends) or held for manual review/edit — the CRM exposes both modes per workflow. - The full ecosystem being pitched: content calendar + article writing agent + growth panel + lead management + email sequencing + knowledge wiki + agent roster, all inside one owned codebase connected by a shared Supabase database and hosted on Cloudflare. - The philosophical frame: "digital sovereignty" — own the infrastructure, own the data, own the agents, don't rent them from GoHighLevel / Notion / a SaaS CRM. Managed Agents is positioned as the piece that finally makes this practical for non-enterprise builders. ## Buildable Ideas - Brand-wiki-backed agent factory — a reusable scaffolding where any new client gets (a) a brand wiki populated from intake, (b) a set of Claude Managed Agents auto-wired to it, (c) a CRM shell exposing the standard tools. Onboarding compresses to under an hour. - Command-centre UI pattern — a single chat panel inside the operator dashboard that speaks to every agent. Replaces jumping between tools. Fits Himiko's desired UX. - Wiki lint agent — a weekly scheduled agent that reads the wiki, checks for stale entries, inconsistencies, missing links, and either fixes or flags them. Directly applicable to Jo's Karpathy-style memory wiki (already seeded). - Growth advisor agent — reads CRM, offers, and wiki; proposes funnels and campaigns; can execute them. Could be ported into Polerie as a packaged consulting tool. - Funnel architect tool — the tool the agent calls to actually build a funnel in the CRM. Generalisable: any deterministic builder (landing page, email sequence, lead magnet) becomes a tool that any agent can invoke. - Agent per CRM verb — client onboarding, lead qualification, email drafting, article writing, pipeline scoring. Each is one agent with one system prompt and one tool cluster. Cheap to add, composable. - Managed Agents migration path for Himiko — Jo's Hermes/Jin/Himiko stack currently rolls its own agent runtime. Claude Managed Agents is now the sanctioned backend and removes that maintenance burden. ## Key Takeaways - Claude Managed Agents is now the default execution layer for any Claude-powered agent in your own product — it removes the need to build agent infrastructure (memory, loop, tool registry) from scratch. - The brand wiki is the critical IP. Without persistent, structured context every agent is generic. With it, every agent is effectively a specialist hire. Jo's memory wiki work is directly on the right path. - A single CRM with a shared Supabase DB + a roster of Managed Agents + a brand wiki replaces GoHighLevel, Notion, ConvertKit, and the typical SaaS stack for a services business. - The UX pattern that makes it usable is the command-centre chat panel — one conversational interface that routes to the correct agent and exposes results in the dashboard. - This video is the direct next step for Atlas — the architecture from part 2 (front end + back end + Supabase + GitHub Actions) is now extended with a managed-agent layer, and the stack is production-ready.