Halo

Most agent debugging tools stop at showing you what happened. HALO is designed to tell you what to change. It takes production execution traces from running agent harnesses, runs them through an RLM-based analysis engine, and produces reports that identify systemic failure patterns — not one-off errors, but recurring breakdowns in planning or execution sequences. The workflow is trace-in, critique-out: your agents run, HALO analyzes the trace corpus, and you get a report you can act on in the next sprint.

The differentiating mechanism is the RLM (Reinforcement Learning from Mistakes) report layer. Rather than simply surfacing where an agent failed, the docs describe an approach that reconstructs the decision path and attributes failures to specific harness components — planning logic, skill invocations, or execution branching. This is the feature that separates HALO from generic observability dashboards: it closes the loop from trace observation back to code change.

HALO fits teams who are past the prototype stage and running agents under real load. It installs via a signed, notarized desktop app on macOS, with a GitHub releases path for other platforms, and a hosted option is available through inference.net for teams who do not want to self-manage. There is no API surface exposed, which means HALO does not slot into an existing pipeline as a callable service — it is a development and optimization tool you run against your traces, not a runtime component. Teams expecting to trigger HALO programmatically as part of a CI step will hit that wall immediately.