What they did

The authors built SimSkill around the Simulation of Urban MObility (SUMO) platform. The agent identifies gaps in its abilities, generates environment-grounded tasks, attempts them using executable tools and code, and checks the resulting artifacts through an action–critic loop.

Successful experience is consolidated into three forms of memory: episodic records of prior attempts, procedural instructions for carrying out workflows, and semantic knowledge about the domain. The system was evaluated on two held-out benchmarks using three backbone LLMs, without modifying their parameters.

Key findings

• SimSkill improved independently verified task completion by up to 25 percentage points.
• The agent accumulated a reusable skill library spanning multiple stages of the traffic-simulation workflow.
• Ablation experiments indicated that procedural and semantic memory made complementary contributions to performance.
• Benefits depended on the backbone model and inference budget; memory neither helped every model nor consistently reduced inference cost.

Why it matters

The study shows how an agent can turn verified interactions with a technical simulator into reusable competence without model retraining. Its combination of natural-language memory with executable code offers a practical design for agents that must preserve and compose skills across repeated, specialized tasks.

Caveats

The reported evidence is confined to SUMO traffic simulation and two held-out benchmarks, so it is unclear how well the approach transfers to other software environments or longer deployment periods. Performance gains are also not universal: they vary with the underlying LLM and available inference budget, and the abstract does not establish whether accumulated memory remains reliable as the library grows.