Why AI Agents Say “Done” When the Work Isn’t | Bounded Loops & graphs | Research

An AI agent can say “done” while its output still fails a check. Bounded Loops puts the completion decision in an independent gate, with limits declared before execution.

Varun Pratap Bhardwaj explains the architecture and shows the CLI, graph monitor, human approval checkpoint, and a real Codex citation-repair example. This is AI Reliability Engineering: define the acceptance condition, run the check, and retain evidence of the result.

Code and examples: https://github.com/qualixar/bounded-loops

Research paper: https://arxiv.org/abs/2609.27871

Qualixar: https://qualixar.com

CHAPTERS

00:00 The completion claim

00:52 Self-reporting and vacuous checks

01:41 Anatomy of a bounded loop

02:29 Spend bounds and global repair

03:25 Terminal, graph and Codex demonstrations

What the demonstrations show:

• Terminal convergence: replay workers, real pytest gates, and ledger verification. The FIXTURE label identifies this example.

• Graph monitor: checks, a join, and a pause awaiting human approval. The downstream publish step has not run.

• Real Codex worker: citation repair checked against a seeded local reference dataset. This run had attempt and time limits, but no token ceiling; it is not evidence of cost savings or verification against an external legal database.

The paper examines gate defects in the authors’ loop catalogue. It does not establish a failure rate for AI agents across the industry. A passing result remains only as reliable as its acceptance gate.

Paper: Bounded Loops: Pre-Run Spend Bounds, Proved Termination, and Verified Completion for Agent Harnesses

Authors: Varun Pratap Bhardwaj, Garima Singh, Arun Pratap Bhardwaj

Engine, catalogue and corpus: Apache-2.0.

SOURCE CLARIFICATIONS

• The paper’s 69-loop catalogue yielded 47 pinned gate-defect regression cases. This is not a 68% industry-wide agent failure rate. The opening “99%” card is unsupported.

• The graph attempt bound is A ≤ (R + 1) Σ mᵥ (Theorem 5.2), under its stated assumptions: a global repair budget, terminating worker/gate calls, and no outstanding human pause. Topological sorting alone does not supply that guarantee.

• The real Codex example had no token ceiling. Its receipt reports 329,462 tokens and a gate check against a seeded local citation dataset.

• The ledger uses a hash chain, with an externally held head needed to detect a rewritten chain; it is not a Merkle-tree proof.

• The engine, catalogue and corpus are Apache-2.0, rather than MIT. The canonical repository is https://github.com/qualixar/bounded-loops.

Production note: Real screen recordings are combined with AI-generated presenter scenes and narration.

#AIAgents #AIReliabilityEngineering #BoundedLoops