What loop engineering is
What is loop engineering?
Loop engineering is the discipline of designing the plan-act-observe cycle an AI agent runs to complete multi-step work: how it decides the next step, calls a tool, checks the result, and knows when to stop. It covers termination conditions, state management, retry behavior, and cost bounds.
What is an agent loop?
The execution cycle a model runs to accomplish a task: plan, call a tool, observe the result, decide the next step, repeat until done. A single chat completion answers a question; a loop does something — reads files, calls APIs, writes results back — by repeating this cycle multiple times.
Why is loop engineering suddenly a named discipline?
Because standing loops — agents that run continuously or across many steps without a human re-prompting each turn — became common in 2026. Standing loops fail in ways single requests don't: infinite retries, tool thrashing, and cost that compounds silently over hours or days. That created demand for a design discipline focused specifically on the loop's internal shape.
Is loop engineering the same as agent orchestration?
They overlap but aren't identical. Agent orchestration usually refers to coordinating multiple agents or sub-agents toward a shared goal — routing, hierarchy, hand-offs. Loop engineering is narrower: it's about the internal cycle of a single agent's execution, whether or not that agent is part of a larger orchestrated system.
Who needs loop engineering?
Any team running agents that take more than one or two steps per task, especially standing loops that operate unattended. Teams using agents only for single-turn Q&A or one-shot code completion need it less — there's no multi-step cycle to design yet.
Designing the loop
What is a termination condition?
An explicit, checkable definition of what 'done' means for a given task — not a vague sense that the work is complete, but a specific state or output the loop can test for. Without one, agents tend to keep refining past the point of diminishing returns.
How many steps should an agent loop run before stopping?
There's no universal number — it depends on task complexity — but every loop should have a hard step ceiling set in advance. A loop that hasn't converged by the ceiling should stop and escalate to a human rather than continuing indefinitely.
What state should carry across steps in a loop?
The minimum needed to make the next decision correctly — often a rolling summary rather than the full history. Carrying the entire conversation forward at every step means token cost grows roughly linearly with the number of steps taken.
How should a loop handle a failed tool call?
Design this explicitly for each tool rather than relying on framework defaults: retry with backoff for transient failures, try an alternative approach for a small number of attempts, then stop and escalate to a human. Indefinite retry on the identical call is the single most common loop failure.
What is tool thrashing?
A pattern where the agent alternates between two or three tool calls without making progress toward the goal — effectively a smaller loop stuck inside the larger one. It's a sign the loop's planning step isn't correctly evaluating whether prior steps moved the task forward.
Cost and tokenomics
Why do standing loops cost more than one-off tasks?
Because every additional step consumes tokens whether or not it advances the task, and standing loops by definition keep running. A loop with a subtle design flaw that would cost a few extra cents in a one-off task can cost thousands of dollars over a week of continuous operation.
What is 'token value per watt per user' and why does it matter for loop design?
It's the metric of how much useful output a loop produces per unit of compute expended — favored over raw accuracy or token count. A well-designed loop that finishes in fewer, more targeted steps scores higher on this measure than one that takes many redundant steps to reach the same result.
Does model routing belong in loop engineering or harness engineering?
Both touch it. Loop engineering decides at which steps routing decisions need to be made (e.g., use a cheap model for simple sub-steps); harness engineering enforces the routing policy and its budget limits across every loop that runs.
How do I measure cost for a loop, not just a single call?
Measure cost per completed workflow, not per request — sum tokens (and compute) across every step of a session and divide by whether the task actually finished. A loop that completes in 8 steps at higher per-step cost can be cheaper overall than one that needs 40 cheap steps to reach the same outcome.
What is a circuit breaker in loop design?
A hard limit — on steps, tokens, wall-clock time, or dollars — that forcibly stops a loop once crossed, regardless of whether the loop believes it's making progress. Circuit breakers are the last line of defense against a design flaw that wasn't caught in testing.
Safety and governance
Does a well-designed loop need a harness too?
Yes. Loop engineering designs how the agent moves step to step; it says nothing about whether each step is safe to execute against real systems. A cleanly designed loop with no harness will still take unsafe actions if nothing is stopping it — the loop just executes efficiently toward whatever it decides to do.
What is silent drift in a standing loop?
A failure mode where a long-running loop keeps operating after its assumptions about the environment stop being true — a service it depends on changed, a data source moved — and nobody notices until the output is visibly wrong. Standing loops need monitoring for this, not just for crashes.
How do I approve high-risk actions inside a loop without stopping the whole loop?
Design an explicit approval checkpoint as a step type: the loop pauses at that step, a human approves or denies, and the loop resumes from the same state. This keeps the loop's design intact while adding a human gate at exactly the steps that need one.
What's the difference between a loop's step ceiling and a token budget?
A step ceiling bounds how many iterations the loop can run, regardless of cost per step. A token budget bounds total spend, regardless of how many steps that took. Both are needed — a loop can hit a low step count and still blow a budget if each step is expensive.
How is an audit trail different for a loop vs a single agent call?
A single call needs one log entry. A loop needs a record of every step — what was planned, what tool was called, what was observed, and why the loop decided to continue or stop — so a human reviewing a bad outcome can reconstruct exactly where the loop diverged from expected behavior.
Relation to other disciplines
How does loop engineering relate to prompt engineering?
Prompt engineering shapes what's sent to the model for a single call inside the loop. Loop engineering decides how many calls happen, in what sequence, and when to stop. A perfectly engineered prompt can still sit inside a badly engineered loop that calls it too many times or never converges.
How does loop engineering relate to context engineering?
Context engineering decides what fills the context window for one turn. Loop engineering decides what state carries across turns in the first place — which determines how much context engineering even has to work with at each step.
How does loop engineering relate to harness engineering?
Loop engineering designs the cycle; harness engineering governs what the cycle is allowed to do. They're addressed together for any workflow that runs unattended, since a badly designed loop and an ungoverned harness fail in overlapping ways — runaway cost and unsafe actions.
Is loop engineering relevant outside of coding agents?
Yes — arguably more so. Operational agents (incident response, Day 2 ops, monitoring loops) are almost always standing loops rather than one-off tasks, which makes termination conditions, retry design, and cost bounds directly relevant to production reliability, not just code quality.
Where do I start if I'm building my first standing loop?
Define the termination condition first, set a hard step ceiling, decide what state carries forward, design failure handling per tool, and wrap the whole thing in a harness before it touches production. The full pre-launch list is in the loop engineering checklist.