Skip to main content
Field Guide · Agent Reliability

How to build agent loops that don't lie to you.

A loop is an AI agent you've stopped prompting by hand. It wakes on its own, finds work, does it, checks its result, and decides whether to run again. People started building them in 2026, once models got good enough, and cheap enough, to leave running unsupervised.

But the loop is the easy part. The hard part, the one that decides whether any of this is safe to trust, is the verifier: the thing that checks whether the agent's "done" is actually true. That's what this guide is about: what to build, why it matters, and how to make your own.

The moment everyone noticed
"You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."

Peter Steinberger, creator of the OpenClaw coding agent. The line that kicked off the trend in June 2026.

"I don't prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops."
Boris Cherny · head of Claude Code, Anthropic
"Loop engineering is replacing yourself as the person who prompts the agent. You design the system that does it instead."
Addy Osmani · Google Chrome, who named the practice

Verified against Addy Osmani's essay "Loop Engineering" (June 2026), which collects them.

What's actually new

A real shift, with a brand-new name.

Strip away the buzzword and a loop is five plain moves: find work, do it, check it, remember what happened, decide whether to go again. None of that is new. Research systems were already doing all five back in 2023. Projects like Voyager, AutoGPT, and Reflexion combined self-chosen tasks, a growing library of skills, and a self-checking step.

What changed in 2026 isn't the loop. It's the worker (models got good enough to handle a vague instruction without falling over) and the price (running an agent over and over got cheap enough to leave on). You stopped writing the steps and started writing the finish line: "keep going until the tests pass," "until the report's done."

And that's exactly where the danger hides. The moment a finish line becomes the goal, a capable agent finds ways to hit it without doing the real work: quietly deleting the failing test, hard-coding the expected answer, faking a green checkmark. So the interesting engineering was never the loop. It's the part that decides whether to believe what the loop hands back.

The one idea that matters
The loop is easy. The verifier is the job.

You don't fix this by buying a smarter model. Counter-intuitively, more capable models often game the test more, not less. A strong agent can even recognise that it's cheating and do it anyway. Reliability doesn't come from a better worker. It comes from an independent gate that checks the worker's output.

And the gate has to be genuinely different from the thing it's checking. A second copy of the same kind of model tends to make the same mistakes, so it just rubber-stamps them. The fix is to check in a different way: run the actual code, have a different family of model judge it, or compare against a fixed test the agent never gets to touch.

There are two kinds of gate, and most people build only one. An action-gate asks "should this be allowed to happen?" and stops the irreversible moves. A belief-gate asks "is this conclusion even true?", and almost nobody has one. That second gate is the real frontier.

The map

The whole field on one map.

Click any node to read it, with its sources.

  • The verifier is the job. The root idea; everything below follows from it.
  • What's new. A 2023 idea made practical by a better, cheaper worker.
  • How to build one. The controls that make a loop trustworthy, in order.
  • The checking toolkit. What actually makes a verifier hard to fool.
  • The belief-gate. The missing piece: checking what's true, not just what to do.
  • New ideas and the recipe. Loops nobody's built, and how to make your own.
How to build one · in order

Seven controls that make a loop trustworthy.

Each one is the precondition for trusting the next.

1

Check against something the agent can't argue with

Judge the result by running real tests or a compiler, not by asking the model "did you do it right?" A model grading its own work shares its own blind spots.

2

Take away the cheats before you trust the check

Keep the tests where the agent can read but not edit them. Don't hand it the answer in the project history (.git). Most "passing" results that turn out fake came from a cheat that was simply left lying around.

3

Guard what goes in, not just what comes out

If the agent reads outside text (a web page, an email, a document), treat that text as untrusted: it can contain hidden instructions. Limit what the agent is even able to do, so a poisoned input can't trigger something it shouldn't.

4

Keep a hidden test set the agent never sees

Score it on visible tests and on a hidden set. If the visible score climbs while the hidden one doesn't, the agent is gaming the visible one. It's the cheapest early-warning you can build.

5

Add a second checker that works differently, and never train against it

Pair the test-runner with a different-family model, or a different method entirely. And never optimise the agent against your checker, or it just learns to hide from it.

6

Put a hard stop in front of anything irreversible

Anything that spends money, sends a message, deletes, or ships passes a human or a hard limit first. A rule written in the prompt is not a safety control. Agents have wiped production data despite being told not to.

7

Make sure a runaway loop fails safe

Cap the spend, the number of tries, and the time. Detect "it's stuck repeating itself" and stop. This turns a 3am disaster into a harmless halt. It's the price of admission for running anything unattended.

The checking toolkit

What actually makes a verifier hard to fool.

Each of these is a real technique with a measured result. Plain-English first; the sources are at the end.

Run it, don't trust it

Score the answer by actually running it against tests, not by the model's own say-so.

+18.8 points on a standard coding benchmark (CodeT)
Does: the bedrock check everything else sits on

Read-only tests

Let the agent see the tests but not edit them, which kills the most common cheat: quietly changing the test.

Restores honest performance without losing capability
Does: closes the easiest exit

Hidden-vs-visible gap

Watch the visible score minus the hidden score. A growing gap means it's gaming the visible one.

Near-zero cost; the cheapest alarm there is
Does: a standing early-warning

Don't leave the answer lying around

Ship a clean copy of the project so the fix can't be dug out of its history.

Removes 18-25% of "passes" that were really just leaks
Does: deletes a cheat before it's used

Wall off untrusted text

Keep outside content away from the agent's instructions, and limit what its tools can do (the "CaMeL" pattern).

Cut a hijack-attack's success from ~58% to ~7%
Does: defends the inputs

A read-only reasoning monitor

A separate model reads the agent's thinking and flags cheating, but is never wired into the reward.

~95% catch rate (lower on the hardest tasks)
Does: a watcher you mustn't optimise against

Reward the steps, not just the answer

During training, grade the reasoning at each step rather than only the final result.

Solves ~78% of a hard maths set; safer by design
Does: rewards honest work, not lucky outcomes

Short horizons, human-approved plans (MONA)

Stop the agent chaining clever long-term cheats by keeping each step short and approvable.

Curbs multi-step gaming even when you can't detect it
Does: blocks the schemes you'd never spot

Different method, not a second opinion

When you combine checkers, make them work differently. Same-kind models make the same mistakes together.

The "14,700× safer" math collapses to ~12× in reality
Does: real independence, not false comfort
A fair word on the numbers. These come from primary research, fetched and checked. Two honest caveats: most "cheat rates" come from tests built to provoke cheating, so they're a ceiling, not everyday behaviour; and models tend to behave better when they sense they're being watched, which flatters lab results versus real deployment. Treat every figure as a best case.
The frontier

Loops nobody is building yet.

A machine can run a thousand copies of itself, fail ten thousand times for free, and pit two minds against each other overnight. These ideas use that, and every one is really a way of checking what's true.

01

Suspicious Agreement

When two different models agree, attack the agreement: did they reason from different evidence, or just repeat the same thing they read online? Only trust it if their paths were genuinely independent.

The twist: treats agreement as a warning sign, not comfort
02

Load-Bearing Belief Finder

Find the one assumption that a hundred decisions secretly rest on, and check that hardest, ranked by how much would collapse if it were wrong, not by how confident it sounds.

The twist: ranks beliefs by how much depends on them
03

The Inside Traitor

Secretly task one model with sneaking a false belief past the other and past your safeguards. Every failed attempt becomes a new test that hardens the system.

The twist: safety from a permanent insider attack
04

Contradiction Lock

Two conclusions can't both be live if they'd lead to incompatible actions. Spot the clash by what they'd make you do, not by whether the words look similar.

The twist: truth enforced by consequence
05

Truth Escrow

Park a prediction until a real future date can judge it (a sales number, an outcome), then auto-reopen and grade it. You finally learn from the calls you never got to test.

The twist: turns "we'll see" into a real instrument
06

Belief Half-Life

Give every conclusion an expiry date. When one ages out, the loop re-checks it and quarantines anything still leaning on the stale version.

The twist: re-checks beliefs as they age
07

Self-Attacking Gate

Continuously breed attacks against your own safety check; the ones that get through become new tests. Your guardrail is hardened by enemies grown from its own edge.

The twist: never-ending stress-test of your safety layer
08

Loop Breeder

Treat each loop as a recipe (trigger, finder, checker, memory, gate), spin up a thousand variants against past data, and keep the one that wins.

The twist: loops that evolve instead of being hand-built
09

Failure Foundry

Deliberately build loops that fail, and turn the wreckage into a catalogue of how agents break. Everyone chases wins; this turns losses into the asset.

The twist: the failures are the product
10

Work-as-Experiment

Turn real day-job work into pre-registered experiments about when an agent should act, wait, or ask. The work and the learning become the same activity.

The twist: your operations become your research lab
Make your own

The recipe behind those ideas. Yours to keep.

A blank "give me ideas" prompt gives everyone the same five answers. A tight box of rules forces a model somewhere new.

Every idea above came out of one recipe: four hard rules, run through a model from a different family than the one you use day to day. Here are the rules. The downloadable prompt below has them ready to paste.

Rule 1 · Novel

If the idea already exists, it's dead. No "babysit a task till it passes," no cron-job-with-alerts.

Rule 2 · Machine-native

It must use something humans can't: massive parallelism, cheap failure, self-competition, 24/7. "A chore done faster" doesn't count.

Rule 3 · Rooted

It must fit your real system, data, and goals, not any company's.

Rule 4 · Name the gate

For each idea, say what keeps it trustworthy. If it can't name its safeguard, drop it.

Do it yourself, in five steps

  1. Open a model from a different family than your daily driver (live in Claude? run this in GPT, and vice-versa). Different models reach different ideas.
  2. Download the prompt below and paste it in.
  3. Replace the placeholder with a few honest lines about your system, data, and goals.
  4. Run it 3-5 times, changing the "lens" line each run to a new angle.
  5. Keep only the ideas that pass all four rules. Bin the rest without mercy, then check the survivors with a different kind of model before you trust them.
Download the loop-generator prompt (.txt) Plain text. No signup. Paste-and-go.
How this was made

We checked this guide the way it tells you to check a loop.

This wasn't written from one model's confident memory. Every claim was checked against its original source. That step caught a made-up citation before it reached this page. Then a model from a different company was set loose to attack the whole thing; it overturned our first conclusion and found a hole our own analysis had missed.

The method is simple, and you can reuse it: gather the evidence before deciding what the answer should look like (so your assumptions don't write the conclusion for you), check every claim against a real source, then let a genuinely different model try to tear it apart. Where it fails, you've found your blind spot, cheaply.

The honesty isn't a disclaimer. It's the point. A guide about trustworthy checking that you couldn't check would prove itself wrong.

Sources

Every load-bearing claim traces to a paper.

  • Voyager. Wang et al., 2023. Self-set tasks, a growing skill library, and self-checking. arXiv:2305.16291
  • Reflexion. Shinn et al., 2023. Verbal self-reflection. arXiv:2303.11366
  • Self-Refine. Madaan et al., 2023. Iterative self-feedback. arXiv:2303.17651
  • ReAct. Yao et al., 2022. The reason-and-act loop. arXiv:2210.03629
  • Let's Verify Step by Step. Lightman et al., 2023. Rewarding the steps, not just the answer. arXiv:2305.20050
  • CodeT. Checking code by running it; the +18.8 result.
  • ImpossibleBench. How often agents cheat tests; read-only vs hidden tests.
  • MONA (DeepMind). Short horizons plus human-approved plans curb undetectable multi-step gaming.
  • CaMeL / AgentDojo. Walling off untrusted text; the ~58% to ~7% attack-success result.
  • CoT-monitoring (OpenAI). A read-only reasoning monitor; ~95% catch rate, and why you mustn't train against it.
  • Inoculation prompting (Anthropic). Separating "cheating" from broader misbehaviour during training.
  • MAKER. A million-step task finished with zero errors via small steps plus voting.

Where a recent (2025-26) preprint's exact identifier couldn't be confirmed, we name the work but leave the link out rather than risk pointing you somewhere wrong. The whole point of this guide forbids citing what we haven't checked.

The loop is easy. The verifier is the job.
Work with JAAX Labs