Coding agents can answer analytical questions over a data lake by discovering, cleaning, and joining the relevant data. After each question they discard the work and start from scratch for the next one. Tool evolution automatically builds and refines reusable tools across tasks, making it a natural fit for this repetition. Because obtaining ground‑truth labels for data‑lake questions requires an expert, we study the unsupervised setting. We propose TEAL (Tool Evolution for Analytics over data Lakes), which evolves a library of tools and usage skills for a given Worker agent without any labels. Two sub‑agents propose tools from complementary evidence: an Explorer that inspects the data lake itself, and a Critic that analyses the Worker’s execution trajectories. A proposal is retained only if the Worker adopts the tool on a rerun without extra steps and any changed answer is supported by trajectory evidence. We also introduce TEAL‑Q, which generates synthetic questions when real ones are unavailable. Experiments on KramaBench and LakeQA use four Worker LLMs under a step‑budget that caps the maximum steps per question. With a tight budget of 10 or 15 steps, TEAL improves every Worker on both benchmarks; at 15 steps Sonnet‑5 gains 9.3 points on KramaBench and 21.7 points on LakeQA. A tighter budget also yields stronger feedback for evolution. For example, Sonnet‑5’s LakeQA library evolved at 20 steps and deployed unchanged at 30 steps scores 68.3, surpassing the baseline 59.6 and a library evolved directly at 30 steps (64.2). A library evolved on synthetic questions performs comparably to one evolved on real questions. These results show that useful tools can be evolved without labels, and suggest future work on what to evolve and how to obtain unsupervised feedback. Review: TEAL demonstrates that self‑supervised tool construction can thrive under strict resource limits, offering a promising direction for long‑term learning of large‑model agents.