What it is
An AI agent is a language model in a loop: it reasons about a goal, calls a tool, reads the result and decides what to do next. Automation is the older, simpler cousin: fixed steps that run without thinking.
The useful question is where to let them run. A common pattern on this site’s own projects is an approval gate: agents research, draft and prepare; anything that cannot be undone waits for a person.
An agent loop with an approval gate
How an AI agent works through a task, and where a person approves anything irreversible.
A diagram of ideas, not a measurement.
Ideas to carry
- 1
Approval gates
Let agents prepare everything, and keep a human on anything irreversible.
- 2
Tools make agents
A model becomes an agent when it can act through tools and see the results.
- 3
Simple beats clever
Plain, composable workflows are often more reliable than elaborate autonomous systems.
Established What we know
- The ReAct method (2022) showed that interleaving reasoning steps with actions improves language-model agents on several tasks.[1]
- Anthropic distinguishes predefined workflows from agents that direct their own process, and recommends starting with the simplest pattern that works.[2]
- Alan Turing’s 1950 paper proposed judging machine intelligence by conversation, the imitation game.[3]
Interpretation The older lens · Greek myth
Homer’s Iliad already has Hephaestus’s golden attendants, made to move and help: the dream of helpful machines is older than engines.
Reading
- 1948CyberneticsNorbert Wiener
- 2022ReActYao et al.
- 2019Human CompatibleStuart Russell
Agent: I booked 40 meetings for you
Me: I asked you to find one email
Connected subjects, projects and writing
- Future possibilitiesSubject · Frontier
Perspective What changes when tools can plan is the central open question of this decade.
- EntrepreneurshipSubject · Creation
Perspective Agents can carry routine operations of a small business; this site’s projects test how far.
- PhilosophySubject · Origin
Perspective Delegating work to machines raises who is responsible for a decision.
- AI and neural networksSubject · Intelligence
How machines learn from data, from 1980s backpropagation to the transformers behind today’s language models.
- Computing and informationSubject · Intelligence
From Turing’s machine and Shannon’s bit to the stack of chips, networks and code underneath everything.
- DEVRIM HoldingsProject · Civilization
The home for the ventures: a small portfolio run by one person, with AI agents doing much of the research, building and monitoring.
- TakeInProject · Creation
Own-brand online ordering for independent takeaway restaurants: orders arrive on the restaurant's own site and print in the kitchen.
- What stays scarce when making things gets cheap?Essay · Civilization
If software, research and design keep getting cheaper, value has to move somewhere. A working list of candidates, none of them assumed safe.
- An operating system for a one-person portfolioExperiment · Intelligence
A control panel where AI agents propose work across several ventures, and anything irreversible waits for a human yes.
- An opportunity engineExperiment · Intelligence
Taking an open-source research framework built for trading and generalising it, so trading becomes one plugin among several.
Sources
- ReAct: Synergizing Reasoning and Acting in Language Models arXiv, 2022. Checked 5 Nov 2026.
- Building effective AI agents Anthropic, 2024. Checked 5 Nov 2026.
- Computing Machinery and Intelligence Mind, 1950. Checked 5 Nov 2026.