Agents · the six steps
Six steps. Every one is a place to stay.
Product names and prices verified August 2026. Re-verify before you rely on a figure.
You already do thisGraduated supervision. An intern does not fail by not being the attending.Here it is called a step. Capability rises. So does the responsibility you hold..
Most people will do excellent work on step two or three and never need six. What rises with each step is not only capability. On step one the vendor holds nearly every risk. On step six, you hold nearly all of it.
Step 01
Conversational specialists
You can describe the work in a paragraph and want a consistent adviser, on demand.
- You provide
- Standing instructions, files, a separate history. You are still in the room.
- Tools
- Grok Workspaces, Claude Projects and Skills, Gemini Gems. Custom GPTs are closed to new individual creation as of August 2026; teach the concept.
- Typical cost
- Free tiers exist. Claude paid plans from about $20 / month.
- Ceiling
- They think with you. Apart from scheduled automations, they act only while you are talking.
- Who holds the risk
- The vendor holds nearly every risk. You hold the charter.
Step 02
Consumer agents that do the work
You have recurring work with files, email, or a browser, and no wish to program.
- You provide
- The agent operates a computer, runs on a schedule, produces finished artifacts. You review checkpoints.
- Tools
- Claude Cowork, ChatGPT Work, Grok Bot (beta), Perplexity Comet.
- Typical cost
- Tied to the host plan. Usage is metered. Grok Bot pricing was still shifting in August 2026.
- Ceiling
- You accept the vendor’s connectors, meter, and where memory lives — unless you externalize the record.
- Who holds the risk
- The vendor still holds the runtime. You hold the gates: what may send, what may leave.
Step 03
Automation platforms
Work already hops between applications, and you think in flows.
- You provide
- A drawn path between your tools, with an intelligent step where judgment is needed.
- Tools
- n8n (self-host possible), Zapier Agents, Make, Lindy, Gumloop, Relay.app.
- Typical cost
- n8n cloud from about $20 / month; self-host has no license. Zapier bills activities. Vague mandates wander, and wandering is billed.
- Ceiling
- You own the flow. The model still sits where you put it — public unless you wire n8n to a local model.
- Who holds the risk
- You hold the flow and the meter. The vendor holds the connectors.
Step 04
Developer’s companions
A terminal does not frighten you, and you want the record to be ordinary files.
- You provide
- An agent that reads and writes folders, runs scheduled tasks, speaks MCP, can spawn focused workers.
- Tools
- Claude Code (recommended workbench), Cursor, Codex, Copilot coding agent.
- Typical cost
- Claude from about $20 / month, or metered API.
- Ceiling
- Full control of the record. Still usually a hosted model unless you point it at local weights.
- Who holds the risk
- You hold the files, the keys, the schedule. The model vendor still sees prompts unless you go local.
Step 05
Frameworks you assemble
You need graphs, crews, or an MCP server of your own, and you have time to maintain them.
- You provide
- LangGraph, CrewAI, OpenAI Agents SDK, a local MCP wrapper around a private model.
- Tools
- Python or TypeScript. Observability via OpenTelemetry into Langfuse or Phoenix.
- Typical cost
- Your time, plus model calls. Self-host the rest.
- Ceiling
- Production control. You become operations.
- Who holds the risk
- You hold almost all of it: code, credentials, uptime, traces.
Step 06
Private, on your hardware
The work cannot leave the room — charts, student records, unpublished files — and you will own the machine.
- You provide
- Weights on disk. LM Studio or Ollama. A steward that stays. No outbound chat.
- Tools
- LM Studio, mlx-lm, Ollama, Unsloth. This is the original HT Sherpa walk.
- Typical cost
- The machine you already own, plus a one-sitting install if you want accompaniment.
- Ceiling
- Memory is the honest limit. You will not put a 70B on 16 GB.
- Who holds the risk
- You hold nearly every risk, and every key. That is the point.