Observe, act, evaluate, repeat: loop engineering as the era after agents
'Every intelligent system in nature runs on a loop — your brain, a thermostat, the stock market. AI for the last three years has been fundamentally non-loopy.'
The definition: loop engineering is 'the discipline of designing systems that prompt AI on your behalf, evaluate results automatically, and improve over time without human triggering at each step.' The one-shot ChatGPT prompt ends at ACT — no evaluation metric, no repeat cycle. The eras stack: chat era (prompt, respond, copy-paste), prompt engineering (better inputs, still one round trip), agentic era (multi-step, tools, some memory — 'but you are still the trigger'), and now loops: 'the machine prompts the AI; you steer.' The driving analogy: 'you don't pedal when you drive a car. You build the engine, set the destination, and steer when needed.'
The industry evidence he assembles: Karpathy's 'I haven't written code by hand' era and AutoResearch — one markdown prompt, ~600 lines of Python, one GPU, hundreds of experiments and optimizations looping toward a goal ('improve the language model's training speed'), agents running '16-17 hours a day'; Claude Code's goal-style command (as-heard 'GOLD') where 'a separate evaluator model checks after every turn... it stops when it passes — you don't watch it work'; and the frontier positioning of Fable/Mythos as models 'built architecturally from the ground up' for days-long asynchronous loops. His summary of Karpathy's phrase: 'expressing your will to agents — express your goals and wishes, and let them do the job once you build that system.'
The goal-command pattern spelled out: define the D2C power-bank brand's competitors, North Star metric, values, and paywall in 15-20 lines of WHAT DONE LOOKS LIKE — then the agent runs turns until the evaluator passes it.
It's the conceptual spine for both sprint days and the Catalyst loop sessions — naming why md files, skills, and crons compose into something better than prompting.
Loop engineering is running the same prompt repeatedly.
Repetition without evaluation is spam. The loop's engine is the evaluate step — named criteria, an evaluator (model or human), and failures that change the next iteration.
You don't pedal when you drive a car. You build the engine, set the destination, and steer when needed.
The fastest progress comes not from better models, but from better loops.
Stop being the one who prompts. Start being the one who designs what prompts.
The car analogy (engine, destination, steering) is the cleanest plain-English definition of your role in this very session's workflow.
Go deeper
In one line: Loop engineering = designing systems where the machine prompts the model, an evaluator scores the result, failures feed back as observations, and the cycle repeats without per-step human triggering; the human designs goals and evaluators, and steers.
'Better loops beat better models' is the operating thesis — progress lives in the cycle design ()
Goal specificity replaces task decomposition: describe the finished product's look, feel, data, and behavior — not the subtasks ()
Model positioning noted as-heard: Fable/Mythos 'designed for long-horizon, days-long asynchronous tasks'... 'of course, it's gone right now for a while' — loop-native models as the direction ()
The AGI framing kept modest: 'give a goal, job done' is the far end of the same axis — 'a hypothesis to keep in working memory' ()
▶ Watch this taught:
Answer from memory first — the recall attempt is what makes it stick. Then reveal.
What turns an agent into a loop?
An EVALUATOR with authority — something that scores the output against the goal and routes failures back as new observations. No evaluator, no loop; just a longer one-shot.
