Track map
Detailed content
🔄 Stop prompting; design loops
The advanced-level shift: stop operating the agent and start designing the system that operates it.
Steinberger: "don't prompt your agents—design loops that prompt your agents". The operator steps out; the system steps in.
It's the mental prerequisite for the entire track; without it, fleets and guardrails are just jargon.
A loop that prompts the agent; the Cherny citation (via third parties).
You replace yourself as the person writing prompts: you design the SYSTEM that does this instead of typing every turn.
Gets you out of the bottleneck; your time goes into design, not every iteration.
Loop engineering; operator × architect.
An agent that INFERS the loops you'd want (from your "vibe") and writes them. The highest rung on the abstraction ladder.
Tempting and risky: abstraction scales up and visibility goes down.
Meta-agent; abstraction ladder.
Osmani: a loop is a recursive goal—you define a purpose and the AI iterates until it completes it.
Every solid loop has 2 pillars: the goal (what) and verification (how it knows when to stop).
Recursive objective; objective + verification.
To run across multiple context windows, externalize the state: git + progress files (Anthropic's harness).
Each session: read the state, build 1 feature, self-verify, commit, leave things clean. The disk becomes long-term memory.
Context window; externalized state; 1 feature per session.
Nate Herk: just because Steinberger does this doesn’t mean it applies to you. There are cadence-based, event-driven, and 24/7 loops.
Not everyone needs 24/7—copying the fleet based on status is a mistake, not engineering.
Cadence × event × 24-7; the right tool for your case.
🧱 Advanced patterns
The fleet pattern kit: Ralph loop, scheduled loops, worktrees, guardrails, adversarial verification, fan-out, and memory.
Huntley/HumanLayer: an infinite shell loop that repeats the same prompt; each cycle mutates the code, which becomes the next starting state.
Powerful and dangerous: “bizarre emergent behavior” if it runs too long. Always combine it with guardrails.
Ralph loop; the one-liner while :; do ...; done.
Lewis (as reported): a job the agent repeats — wakes up on a schedule/event, reads the state, does 1 scoped job, writes it down, sleeps.
The manager classifies the backlog; the worker implements tasks and opens PRs; Issues as the control plane; a human approves the merge.
Split manager/worker; control plane; fixed permissions.
A worktree isolates parallel work; a sandbox isolates the environment; fixed permissions isolate what the agent CAN do.
Fixed permissions are how you let a loop run without doing something irreversible.
Git worktree; sandbox; fixed permissions.
Grassroots consensus: "the guardrails are the real work, not the agents." The worker opens a PR but never merges it; a human approves.
The agent is the easy part; the hard part is designing the boundaries around it.
Guardrails; permissions prompt; final human gate.
A second agent (or panel) that tries to BREAK the work, prompted to refute it. It approves only when it can’t.
The person doing the work never grades their own work—verification taken to the extreme.
Verification specialist (speculation, not fact); refutation.
A manager prompting agents that prompt agents ("Russian nesting dolls"). The pattern: orchestrate → delegate → verify.
Fan-out (A→N) adds speed; verification (N→1) prevents N errors from turning into a broken product.
Fan-out; orchestration; A → N → 1.
The loop spans episodes: act → feedback → reflect in text → save to episodic memory → do better next time.
Learns through language, not retraining. A worker that retains what it learned improves with every ticket.
Cross-trial memory; Reflexion; episodic memory.
🛰️ Risks, cost, and case study
The course wrap-up: the real dangers of the autonomous loop, and how this very course was audited by a maker/checker loop.
Osmani: hand over full control to the loop and stop reading the output. You remain the reviewer, especially at first.
The loop can’t sense when it’s drifting; you’re the missing sensor.
Cognitive surrender; delegate with review × surrender.
Firecrawl warns about cost and "comprehension debt." Sullivan rejects "tokenmaxxing" and "promptchud," too.
The middle ground: agents that reproduce real problems in your product.
Comprehension debt; tokenmaxxing; promptchud.
Nate: there have been 12h+ loops that weren’t useful. Useful ones usually run ~35 min to a few hours.
A "chunky loop before bed" (4–8h) can deliver; but you don’t need a 4-day loop.
Duration scale; chunky overnight loop; drift.
AI never gets it right on the first try; delegating the feedback-and-iteration cycle to AI makes quality improve much sooner.
It's why the risks are worth the cost: the curve reaches “good enough” sooner.
Human line (slow) × loop line (fast) up to 90–95%.
This course came out of a maker/checker loop: DISCOVER → EXTRACT (maker) → VERIFY (separate, adversarial checker) → SYNTHESIZE.
The model that FINDS the source never EVALUATES it — separating the roles is what builds trust.
Pipeline discover→extract→verify→synthesize.
Thumbnails (rubric + scorer), 3JS plane (render in the browser, ~37 min), Abbey Road in CSS (score ≥9 or cap at 8, stopped at V7), video editing (beats aligned).
In all of them, the bottleneck is the same: making verification objective enough to know when to stop.
Concrete stop condition; dedicated scorer; pass cap.
Use a loop for repetitive, reviewable, high-value tasks with a built-in checker. For everything else, prompt once.
Run loops at the right cadence for YOUR use case — not 24/7 for status. The course’s phrase: loop only with verification.
“Do this / Skip this”; loop only with verification.