Brain-and-Hands Agent Architecture
Pair a reasoning model with an agent that can act in real systems.
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 5
- Confidence
- 94%
The Brain-and-Hands Agent Architecture models an operational AI system as two complementary layers. The brain is the reasoning model that interprets intent, decomposes an outcome, and chooses actions. The hands are the agent runtime, integrations, and tools that execute those actions in external systems. Neither layer is sufficient alone: reasoning without execution leaves humans to perform the work, while execution without sound reasoning creates brittle automation. Designers therefore select the reasoning layer, expose a narrowly scoped set of capabilities, define how planned actions become tool calls, and verify results from the affected systems. Security controls belong at the boundary between brain and hands because that is where generated intent becomes a consequential action. The architecture applies to personal assistants, marketing operations, sales workflows, support systems, and software agents.
Origin
Extracted from Marketing Against The Grain through Eric Siu's description of the model as the brain and OpenClaw as the hands.
Core principles
- 01Separate reasoning from real-world execution.
- 02Connect intelligence to tools that can change system state.
- 03Treat permissions and integrations as part of the architecture.
- 04Evaluate the combined system by completed work, not generated text.
How to run it
- 1
Choose the brain
Select a reasoning model suited to the workflow's complexity, context, and reliability requirements.
Pro tip Evaluate it using realistic decisions from the intended workflow rather than general benchmarks alone.
Watch out A capable model cannot compensate for missing or unsafe execution controls.
- 2
Define the hands
List the concrete tools and integrations required to perform the work, such as email, calendars, browsers, or product APIs.
Pro tip Map each capability to a specific business outcome.
Watch out Avoid exposing broad system access when a narrow API can perform the same function.
- 3
Design the action handoff
Specify how the model selects tools, supplies arguments, interprets results, and decides whether another action is necessary.
Pro tip Use structured tool contracts and explicit success criteria.
Watch out Ambiguous tool interfaces can cause correct reasoning to produce incorrect actions.
- 4
Constrain execution
Apply least privilege, approval gates, logging, and limits at the point where reasoning becomes action.
Pro tip Require confirmation for irreversible or externally visible actions.
Watch out Prompt instructions alone are not a sufficient security boundary.
- 5
Verify end to end
Run a bounded task and confirm the intended state change in the target system, not merely in the agent's response.
Pro tip Include failure and partial-completion cases in testing.
Watch out An agent claiming success does not prove that the external action succeeded.
In the wild
A reasoning model examines site data, identifies likely SEO issues, and decides what evidence is needed. The agent layer logs into an approved product, gathers the relevant records, produces content or tickets, and flags issues through the team's normal workflow.
→ Reasoning and execution combine into an operational assistant rather than a recommendation-only chatbot.
An agency agent interprets a client's request to add a menu, determines the required site change, uses its authorized website tools to make the update, verifies the result, and tells the client it is complete.
→ The system completes a real client-service workflow across communication and website tooling.
Common mistakes
Building a brain without hands
A model that only drafts instructions still leaves the human responsible for every consequential action.
Giving the hands excessive reach
Connecting email, files, browsers, and credentials without strict scoping creates a large compromise surface.
Trusting self-reported completion
The architecture must inspect tool results or external state rather than accepting the model's assertion that work is done.
Is it for you?
Best for
Teams building assistants that must perform multi-step work across messaging, files, browsers, calendars, or business applications.
Not ideal for
Simple question-answering tasks where generated information is sufficient and no external action is needed.
From the transcript
“So he describes it here clawed code the brain, which is what OpenClaw is using, and OpenClaw the Hands. It's able to actually do the…”
“It's a true autonomous personal assistant.”
“A true autonomous agent doing employee work.”
From the episode
770,000 Agents, 0 Humans: Inside the First AI Social Network