Should you evaluate it?
Evaluate Microsoft Agent Lightning when preserving an existing agent's tools and execution flow is a training requirement and your team can validate the GPU environment. For a training job centered on a model and dataset, compare the requirement with TRL first.
Agent Lightning separates agent execution from policy training through a trainer, API gateway and rollout controller. The README at the measured commit describes a trainer running Verl and vLLM, a gateway capturing model requests, and a controller launching agents locally or as Kubernetes Jobs. Those are documented capabilities; our sandbox did not reproduce a training result.
What the sandbox established
Our run took place on September 18, 2026, at commit ff94575, full SHA ff9457587fb6ec900e16e93be9ad2d77409afa08. Installation recorded 83 packages occupying 203 MB and took 20 seconds. The build command returned zero in 1 second. The test command exited 1 after 8 seconds.
The recorded parser summary contains 82 passed, 0 failed, 2 skipped and 2 errors, with a stored total of 84. The supplied record does not verify exhaustive unique test counts, so the stored total cannot support a passing-test percentage. No new execution or complete-log reconciliation was performed for this correction.
The log tail identifies collection errors in tests/verl/test_agl_rollout_manager.py and tests/verl/test_rollout_level_advantage.py. Both report ModuleNotFoundError: No module named 'numpy'. This identifies a dependency problem to investigate without establishing a defect in the training algorithm. The pytest footer reports 4.42 seconds; the harness records 8 seconds for the command. These are separate timing records.
Pip-audit recorded one advisory and one affected package. The supplied record does not identify the advisory, its severity or the number of packages scanned. Scan coverage is unknown, and this result is not a security assessment.
What the GPU instructions require
The supplied installation guide supports CUDA 12.9 or 13.0. It describes Verl 0.7.1 with vLLM 0.12.0, or Verl 0.8.0 with vLLM 0.20.2, with FlashAttention 2.8.3 built locally. It also says tasks upload logs and trajectories to Weights & Biases by default. Review that destination before using sensitive trajectories.
The supplied quick start requires one machine with one A100 GPU. The quick-start guide also requires the Calc-X dataset and additional example dependencies. Its launcher starts Ray, the model backend, gateway and local controller before running training. That local runner is unsupported on native Windows; the guide directs Windows users to a Linux environment such as WSL.
Both supplied setup extracts display a development-documentation warning despite their stable URLs. Their instructions are not verified compatibility evidence for the exact measured commit. The sandbox record does not establish a completed Calc-X rollout, GPU inference, policy update, Kubernetes Job or improvement on a held-out benchmark.
Compare the training layer you need
TRL provides trainers for supervised fine-tuning, GRPO and DPO. Its README describes training language models or adapters on a custom dataset. Evaluate that approach when your job centers on a model and dataset. The supplied evidence does not establish TRL as a drop-in replacement for Agent Lightning's existing-agent integration, and we did not compare setup time or training performance.
Verl is Agent Lightning's documented training backend. Its README describes RL dataflows and modular integration with training and inference frameworks. Direct integration is an engineering option for teams intending to own that work. Treat Verl as a component of the documented Agent Lightning stack, rather than assuming it replaces the gateway and rollout controller.
What to verify next
Reproduce the two collection errors in your intended environment and check the missing dependency against the installation instructions. Obtain the recorded audit advisory before judging its significance.
Then validate a small Calc-X workload through agent execution, captured interactions, reward calculation and a policy update. Check cleanup and logging destinations during the trial. Use a held-out evaluation to decide whether training improves your agent. These are proposed evaluation steps, not results from our sandbox.
Editorial assessments
Setup ease is assessed at 2/5 because the documented training route requires a coupled GPU environment and a local FlashAttention build that our run did not validate. Maturity is assessed at 3/5 because the supplied architecture and examples provide a basis for evaluation, while the nonzero test outcome and missing workload validation leave adoption questions unresolved. These are editorial judgments, not performance measurements.
Documentation and community are unassessed. The measured-commit README lists installation, quick start, trainer, gateway, controller and asynchronous-training guides plus examples. Only the supplied README and setup extracts were inspected; a guide index does not establish documentation quality across those topics. The README also lists DeepWerewolf, AgentFlow and Youtu-Agent and provides contribution routes. Those upstream listings do not establish adoption scale, community quality or support responsiveness.