The Reality Loop

AI agents can do in minutes what used to take days, but they are temporary reasoning processes: broadly capable, forgetful, and probabilistic. The unit of complex work is therefore not the task. It is the verified iteration.

Old: “Go do this thing.”
New: Act → observe → prove.

The loop

flowchart TD
    O["ORIENT<br/>What is observably true now?"] --> C["CHOOSE<br/>Smallest, most informative act"]
    C --> A["ACT<br/>Bounded intervention"]
    A --> B["OBSERVE<br/>Logs, diffs, tests — not vibes"]
    B --> V["VERIFY<br/>Acceptance condition met?"]
    V --> U["UPDATE<br/>Record what was learned"]
    U --> O

Each pass must produce:

Artifact	Evidence	State update	Next decision
What changed	What proves or falsifies the claim	What the durable record now says	Continue, revise, stop, or escalate

State lives outside the agent

Context is compressed, memory decays, and reasoning varies between runs. Anything that matters—specs, decisions and why they were made, tests, definitions of done—must live in files, repositories, and trackers that survive the process that produced them.

Fresh-agent test: could a new person or agent reconstruct the work from what is written down? If not, the system is fragile.

Decompose by uncertainty

Split work around the riskiest premise, not merely by function. Test the assumption that could invalidate the system before turning it into infrastructure.

A good work unit has:

• One objective
• Limited context
• An inspectable output
• A stopping condition
• A way to verify it

Claims attach to evidence

Language should compress evidence, not substitute for it. “It works” is rejected without an anchor. “It passed these three cases and still fails in this documented condition” is useful.

Creation and judgment stay distinct

Generation is cheap; selection is the bottleneck.

flowchart LR
    E["Explorer<br/>proposes"] --> B["Builder<br/>makes"]
    B --> C["Critic<br/>attacks"]
    C --> J["Judge<br/>decides"]
    J --> D{"Continue,<br/>revise, or stop"}

Different people, agents, prompts, or automated checks may fill these roles. What matters is that creation and judgment do not collapse into one unchallenged process.

Where humans go

Human judgment has the greatest leverage in purpose, ambiguity, values, and irreversible risk. Escalate deliberately when:

• The goal is unclear
• Evidence conflicts
• The action cannot be undone cheaply
• Repeated cycles are not reducing uncertainty

Autonomy is earned

Level	Permission
0	Suggest
1	Prepare
2	Act and report
3	Act within explicit bounds
4	Manage the loop
5	Manage nested loops

The reason to grant more autonomy is not a smarter model. It is an environment where failure is detectable and recoverable.

The bet

The people who get the most out of agents will not be those who ask them to do the most, but those who build systems where agents can be wrong cheaply, learn quickly, and become reliable.

Not autonomous execution. Reality-guided agency.

Read the original manifesto