This is a stack reveal. No affiliate links, no sponsored mentions. What follows is exactly what Avakata runs as a one-person operation — the models, the orchestration, the evaluation gates, and the delivery infrastructure — and why each piece is there.
The stack is not aspirational. It is what is running in production today.
The four layers of the stack
Every output Avakata ships passes through four layers:
- Generation — the AI models that produce content, copy, and analysis.
- Orchestration — the layer that wires agents together, manages triggers, and routes work.
- Evaluation — the critic agents that check output before it ships.
- Delivery — the CMS and site infrastructure that publishes and measures.
Each layer has a distinct job. Conflating them is the most common mistake in agentic system design.
Generation layer: Claude as the primary model
Three models are in active use. Each has a defined role.
Claude handles long-form content and brand voice work. It is the primary generation model for Field Notes posts, client deliverables, and anything where tone consistency matters. Claude follows negative constraints — the banned words, the structural rules, the voice guardrails — more reliably than GPT-4. It also produces fewer hallucinations on factual claims, which matters when outputs go to clients without a human fact-check layer.
ChatGPT handles structured data tasks and code. Extraction, transformation, schema generation, and scripting. It is faster and more predictable on well-defined tasks with clear input/output contracts.
Gemini handles research synthesis. When an agent needs to pull signal from a large corpus and compress it into a usable brief, Gemini is the model in that slot.
Model selection is not loyalty. It is task routing. The orchestration layer decides which model gets which job.
Orchestration layer: the proprietary graph
The orchestration layer is custom-built and stays under NDA. It is not a wrapper around an existing framework.
What we can say: it manages 160+ specialist agents, routes work based on trigger conditions, and maintains a memory layer of client context that persists across sessions. Agents do not start from scratch on each run. They inherit relevant state.
The orchestration layer is the product. The models are commodities. Any sufficiently motivated team can access Claude, ChatGPT, and Gemini. The durable advantage is in how work is decomposed, routed, evaluated, and looped back.
This is also why the orchestration layer is not discussed in detail here. The architecture is the moat.
Evaluation layer: the critic gate
Every output passes through a critic agent before it ships. No exceptions.
The critic checks four things: brand voice compliance, GEO structure (is the content citable by AI answer engines?), factual accuracy against source material, and conversion principles. It does not check for grammar. That is a lower-order concern.
Rejection rate is roughly 23%. That means roughly one in four outputs goes back to the generator with specific failure reasons attached. The generator does not get a vague "try again" — it gets a structured failure report: which rule was violated, where, and what the expected output looks like.
The 23% rejection rate is not a problem. It is the system working. Output that clears the critic gate is materially better than output that does not.
Delivery layer: Sanity CMS + custom site
Sanity handles content management. The custom site handles delivery and measurement.
GA4 provides traffic signals. Hotjar provides behavioral data — scroll depth, click patterns, session recordings on key pages. Search Console provides GEO citation tracking: which queries surface Avakata content in AI-generated answers, and how that changes over time.
The delivery layer is not passive. It feeds signals back to the orchestration layer. Traffic patterns, engagement data, and citation signals inform what gets produced next and how. The loop is closed by design, not by manual review.
Weekly time investment
The honest number is 8 hours per week of human time.
- Content review and approval: 2 hours
- Client memos and strategy: 3 hours
- Agent monitoring and refinement: 1 hour
- Business development: 2 hours
The agents handle the rest. Content production, research, tagging, SEO structuring, FAQ generation, internal linking — all agent-run.
The 8-hour figure is not a marketing claim. It is what the time logs show. The constraint is not capacity; it is the quality of the decisions made in those 8 hours.
What we do not run
Absence is as informative as presence.
No Zapier. Replaced by custom orchestration. Zapier is adequate for linear automation. It is not adequate for conditional, stateful, multi-agent workflows.
No Notion AI. Replaced by dedicated agents. Notion AI is a general-purpose assistant bolted onto a note-taking tool. Dedicated agents with defined roles and memory layers outperform it on every task we tested.
No general-purpose AI assistant for business tasks. Replaced by specialist agents. A single assistant that does everything does nothing particularly well. Specialist agents with narrow mandates produce better outputs and are easier to evaluate.
Fewer tools, more depth. Each tool in the stack has a specific job and is accountable to a measurable output.
We send a monthly stack update — what changed, what we added, what we cut — to Field Notes subscribers. Get it at avakata.agency/contact.html.
If you want to understand whether an agentic setup makes sense for your operation, a discovery call is the right next step. No pitch deck — just a direct conversation about your current stack and where the leverage is. Book at avakata.agency/contact.html.