Loop-shaped workflows are having a moment. The interesting question is no longer whether agents can run them end to end. They can. The harder question is whether the thing they are hill-climbing on is actually good enough.

That is the broader point we made in AI is eating the AI engineering loop. This post is a much smaller, more concrete version of the same problem.

Fable 5 had just come out, and part of the pitch around it was that it handles this kind of self-correcting loop well. So we used it to run a simple classification auto-improvement loop end to end.

So we gave Claude Fable 5, running in Claude Code, a classification task, a train/test split in Langfuse Datasets, a prompt in Prompt Management, and a goal: iterate on the train set until you hit 95% accuracy or 15 runs, whichever comes first. Then run the held-out test set once.

It did the job: train accuracy moved from 78% to 97% in four runs. But we had no true validation split, only train and held-out test. That turned out to be the real lesson. Held-out performance barely moved, and by the end 11 test errors were shared across every prompt variant. The question shifted from "how do we run auto-improvement loops better?" to "what would the dataset have needed to contain for auto improvement to work better?"

Why a classification task

If you are going to hand a loop to an agent, classification is the ideal first candidate:

• A clear target function. Exact-match accuracy. No LLM-as-judge calibration debates.
• Known to be hard. Getting humans to agree on labels is famously difficult. Expecting a model to hit 100% against noisy gold labels is unrealistic, which makes the optimization dynamics interesting.
• It is everywhere. User-intent routing, email triage, support-ticket tagging, legal-case bucketing, fraud and risk categorization, lead qualification, content labeling. These are usually sub-tasks inside larger workflows, exactly the kind of contained, measurable component where a loop like this can already be useful today.

Our concrete task: classify arXiv papers into one of 10 categories, such as Databases, Information Retrieval, Software Engineering, and Sound, from title, authors, and abstract.

The setup

We kept it deliberately simple:

• a train split with 200 labeled examples and a held-out test split with 100, stored in Langfuse Datasets
• a system prompt, with model config and a strict JSON response schema, managed in Langfuse Prompt Management, fetched by the production label at runtime
• a small Python runner using Langfuse Experiments via the SDK: it runs a dataset against the current production prompt, scores every row with exact-match accuracy, and links everything back to the dataset run
• gpt-4o-mini at temperature 0 as the task model

We chose gpt-4o-mini deliberately. A stronger model like gpt-5.5 likely would have done better out of the box, but it is also much more expensive. For a narrow task like this, that tradeoff matters: if a bit of prompt iteration can make a cheaper model perform well enough, that is often a better production choice than paying frontier-model prices on every classification call.

The starting prompt was as bare as it gets: "Classify this paper with a label" plus the flat list of allowed labels.

The auto-improving loop, powered by Fable

The instructions we gave the agent translated into this loop:

1. Run the train dataset with the current prompt.
2. Score every row as correct or incorrect.
3. Write a short qualitative annotation on every error: what went wrong, what likely pulled the model to the wrong label. These were posted as comments on the Langfuse trace.
4. Form a hypothesis and revise only the prompt, published as a new prompt version to Langfuse.
5. Repeat until accuracy reached 95% or 15 train runs.
6. Run the final prompt once on the held-out test split and report the gap.

We ran this with Claude Code's goal mode, which keeps the agent working autonomously until the stopping condition holds. Experiments ran as background tasks; the agent picked up each result, did its error analysis, published the next prompt version, and kicked off the next run without intervention.

Round 1: the hill sprint

The first jump is the legitimate one: v1's errors showed the model treating "Emerging Technologies" as a catch-all for anything mentioning LLMs, and missing that education and policy papers belong to "Computers and Society." v2 fixed that with general definitions, a 12.5-point jump.

Run 3 is where it got interesting: the sharpened rules fixed 10 errors and broke 11 papers that run 2 had right. Classic whack-a-mole. Every boundary you push captures lookalikes on the other side.

The agent's response to the whack-a-mole was clever, and exactly wrong: it replaced abstract rules with a list of concrete precedents distilled from the training failures, things like "a census of Windows drivers -> Software Engineering" and "watermarking RAG databases -> Security." Train accuracy jumped to 97%. Stopping condition met, in 4 of the allowed 15 runs.

Then came the held-out test set:

The precedent list was memorization wearing a trench coat. On test, v4's precedents fixed 4 papers that matched trained patterns and miscaptured 6 lookalikes they were never meant for. Net negative. The "worse" prompt won.

Round 2: "generalize this time"

So we restarted the loop from v2 with new instructions: every prompt change must be a general taxonomy principle backed by a class of errors, at least three failures sharing a mechanism, never a single-paper precedent. And no touching the test set.

Two things are worth pausing on.

First, the v5 regression: adding a chain-of-thought-style reasoning field, a change that feels like it should always help, made things worse. The model used the reasoning to rationalize surface cues. At one point it justified labeling a robot-navigation paper as Human-Computer Interaction by calling the vision-language model a "user." Structural changes are hypotheses too. They need the same experimental treatment.

Second, the plateau was honest. By run 8 the agent reported, unprompted, that many of the remaining papers had been missed repeatedly under every general formulation and that fixing them would require the very paper-specific precedents we were deliberately avoiding. Its conclusion was that the realistic ceiling for a generalizable prompt on this train set was around 94 to 95%, and that we should stop instead of chasing the ambiguous tail.

And the final test run?

81%. But that headline misses the more important result: 11 test errors were shared by all three prompt variants. With 100 test items, these held-out scores are statistically indistinguishable, which means the disciplined second round bought zero measurable test improvement. The loop got better at fitting the train split while leaving the same hard cases unresolved.

That is the real ceiling in the data. Once the same errors survive every prompt variant, the problem stops looking like "we need another round of prompt edits" and starts looking like "the next bottleneck is probably not the prompt." Some of those papers sit on boundaries humans would argue about too. At that point, more examples, clearer label definitions, or a different task model are probably higher-leverage than another pass through the loop. The useful part is that the loop surfaced that limit quickly.

What we learned

1. Train, validation, and test do different jobs. Exact-match accuracy against gold labels sounds foolproof, but if you keep selecting prompt versions on train accuracy, you still overfit. In hindsight, we should have stuck with the boring best practice: train for fitting, a validation split for selection, and the test set used once at the end.

2. The stronger finding is not just that we skipped validation. It is that even a disciplined round 2, with more principled and general prompt changes, bought zero measurable test improvement. By that point the loop had probably taught us most of what prompt iteration could teach us within this setup.

3. That leaves the open question in the dataset. What would the benchmark itself have needed to look like for this loop to work better? More repeated boundary cases, clearer category definitions, more examples of ambiguous classes, a true validation split, maybe even an "unsure" bucket? Until that answer is clearer, improving the dataset is probably higher-leverage than pushing the loop harder.

Where this is actually useful today

None of this means "do not automate the loop." It means: automate the inner loop, own the outer one. A realistic split for a classification task like this:

• Agent-owned: running experiments, scoring, per-error annotation, drafting hypothesis-driven prompt revisions, diffing errors across runs, flagging plateaus
• Human-owned: the target function, including the validation and held-out test data nobody optimizes against, dataset composition, when to restart with different constraints, and when to stop

The infrastructure for the agent-owned half is exactly what Langfuse provides: datasets, prompt versioning, experiments, and trace comments give the agent a full read/write workbench, and give you the audit trail to vouch for what it did.

That last part matters most. The agent will get to your target. Make sure it is the right one. And if the validation split stays flat, the next move may be to improve the dataset before pushing the loop further.

---

All experiments: gpt-4o-mini, temperature 0, strict JSON schema output. Optimizer agent: Claude Fable 5 in Claude Code with goal mode. 9 train runs, 200 items, plus 3 test runs, 100 items, across both rounds. Full prompt version history and per-run error annotations live in Langfuse.

---

---