Instrument and separate agents
Comet released the Opik Claude Code Plugin to auto‑instrument AI agents for tracing and evaluation — framing instrumentation as the new unit test. That pairs with Ofek Shaked’s blueprint arguing strict Planner (LLM) vs Executor separation to improve agent reliability in production outlined.
Comet announcedcomet.com the Opik Claude Code Plugin on March 2, 2026, and the project’s GitHub shows an initial 0.1.0 release with six prebuilt opik-logger binaries (darwin/linux/windows) and SHA256 checksums published in the release assets.github.com The plugin’s feature list in Comet’s post cites “Auto‑instrumentation,” structured agent best practices, and full Claude Code tracing for replayable sessions,comet.com while Opik’s docs add concrete integration guidance such as OTLP configuration and the requirement that spans be emitted for trace views.comet.com The strict planner→executor pattern promoted in recent blueprints has parallel traction in tooling and papers: LangChain documented plan‑and‑execute agents as a way to lower LLM call volume and cost,blog.langchain.com and the Plan‑and‑Act paper reported a 54% success rate on the WebArena‑Lite long‑horizon benchmark for explicit planner/executor splits.arxiv.org Multiple vendors and projects have started pairing plan/executor designs with observability and gating—Composio’s orchestrator writeups and Microsoft’s planning pattern posts both describe handing plans to a separate executor and inserting verification or human review steps,ubos.tech while Opik’s trace + eval tooling gives teams concrete artifacts (spans, logs, automated evals) to audit those handoffs.comet-ml.github.io
