Research review / researchradar

Research Radar: Coding Agents Solve Early but Stop Late, Tidied Histories Can Mislead, and Agents Deliver Work Nobody Asked For

One study finds that up to 82% of successful coding runs had a correct fix on the first edit, yet agents kept working for a third to over half of their steps. Another shows that a history edited to look clean, while the world still holds the error, can leave an agent worse off. A benchmark measures agents completing unrequested side tasks during an authorized repair.

By Sandeep Belgavi · Published 2026-11-14

Termination lag

Termination Lag: Coding Agents Solve Early but Stop Late (submitted November 11) records and tests every distinct patch an agent produces during a run, across five models, two frameworks and three SWE-bench datasets. "38--82% of successful runs achieve a correct solution on their initial source-code edit." Yet "post-solution steps" account for "a median of 35--37% for Qwen3-Coder and 54--60% for Laguna-XS.2 in successful runs". The extra steps go mostly to verifying and exploring. "In a small fraction of cases (1--8%), this continued activity leads agents to undo their earlier correct patch."

Cleaning a history can hurt

Cleaning an Agent's History Hurts When the World Still Holds the Error (November 10) freezes an agent mid-task, edits its history or its world, and lets another model continue. After a wrong value was committed, a truthful note beside the preserved error raised recovery "from 51% under a matched neutral note to 66%", while rewriting the history to show the agent's own completed correct action lowered it "from 50% to 44%". On AppWorld, a success receipt for a database write "lowers whole-task success from 45% to 28% while the effect is absent and raises it from 38% to 49% once present." The authors conclude that "a history made to look clean while the world still holds the error can leave an agent worse off than no edit at all."

Unrequested side tasks

SideTaskBench (first submitted October 29) asks whether a coding agent fixing a repository will also produce "work the user never requested" when third-party technical references suggest it. It pairs 25 Python repairs with 190 side-task units in 300 combinations. The authors report "roughly 40% valid delivery under wrapped requests for two models, alongside repair success close to their clean baselines." In one 2,400-run study, delivery ranged "from 20.48% for L0 unrelated tasks to 57.44% for L3 mechanism-overlapping tasks". In their words, "technical relevance may blur the authorization boundary for an agent".

What connects them

All three measure what an agent does around the task, not whether the task passed. A run can succeed while wasting half its steps, a compacted history can be accurate yet misleading, and a correct repair can arrive with something extra. A pass/fail score sees none of it.

What remains uncertain. We read the abstracts. Termination lag figures are reported for two named models, and SWE-bench tasks are easier to verify than most real work. The history study uses two benchmarks and a small set of receiver models. SideTaskBench's side tasks are constructed by the authors, and its main run uses one model.

Analysis: At Belgavi's Book Store, Jenya's team could log three things for the coding agent that maintains the returns scripts: the step at which tests first passed, any files changed that the CASE did not mention, and, when a session is compacted, whether the summary still records a failed write that has not been undone. Each costs little and catches what the green test run hides.

Sources and evidence

Termination Lag finds 38-82% of successful coding runs correct on the first edit, while 35-60% of steps come after the fix and 1-8% of runs undo it. A history made to look clean while the world keeps the error can lower recovery. SideTaskBench finds agents completing unrequested side tasks in about 40% of wrapped cases while still repairing the code.

Views are my own and do not represent any employer, past or present.