IBM Research — Why Your AI Agent Might Not Ace the Same Task Twice

An agent that succeeds 77.4% of the time on average may succeed every time on only about half its tasks. IBM Research measured this hidden flakiness and halved it, without retraining the model.

Key takeaways:

• Averages hide flakiness. On AppWorld, a GPT-4.1 ReAct agent scored 77.4% Mean@5 but only 53.0% Pass^5 (all five runs succeed), a 24.4-point consistency gap.
• Find where it hesitates. A black-box Consistency Analyzer resamples the model at each step of a trace. Where the samples disagree, it flags a decision the agent is guessing at.
• Patch only the weak spots. Each flagged step becomes a targeted consistency guideline that ALTK-Evolve retrieves and injects when the agent reaches a similar state.
• Results. Pass^5 rose from 53.0% to 69.0% and the gap shrank from 24.4 to 12.0 points, with no loss in average accuracy. Medium tasks gained 22.9 points and hard tasks 14.3.
• The lesson. Report consistency alongside averages before trusting an agent with critical work.

Chapters:

00:00 Would you trust a flaky agent?

01:32 How benchmarks measure success

03:03 Same average, different reliability

04:35 Why flakiness matters in business

06:05 Fixing without retraining

07:37 Resampling to find hesitation

10:41 Generating consistency guidelines

11:26 Retrieving guidelines at runtime

12:59 Pass^5: the strict metric

14:34 Results

16:53 Related studies

18:24 Better telemetry

19:55 Recap

Sources:

• Your Agent Aced the Task. Will It Do It Again? (IBM Research, Hugging Face blog, Sep 2026)

Related episode: E27: Grab — Why Public Benchmarks Miss Production AI Failures

Enjoyed it? Follow Data Science in the Wild (one real production ML/AI system or paper per episode, hosted by Peter Liu) and rate it on Apple Podcasts or Spotify.

Also from Peter: The GenAI Evolution Atlas, Wednesday AI deep dives and Friday news roundups.

Made with AI-generated voices; curated by Peter Liu from the sources above.