EVALS

Cursor's latest agent-swarm experiment rebuilt a SQLite-like database from documentation and passed a held-out conformance suite. The bigger story is the shift from assigning code tasks to specifying, measuring, and governing a goal.

A July 2026 paper from Tencent Hunyuan turns agent harnesses into behavior-level maps. The useful lesson for builders is simple: code search is not enough when one behavior spans prompts, tools, state, permissions, and runtime policy.

SkillHone is a July 2026 paper about evolving agent skills across sessions. The useful takeaway for developers is simple: do not save only the latest SKILL.md. Save the decisions that explain why it changed.

StateM pushes Terminal-Bench 2.1 to 95.3% raw accuracy by scaling the harness around the model. The lesson for coding-agent teams is that runbooks, state, and recovery loops now matter as much as model choice.

A fair comparison of Braintrust, Langfuse evals, Promptfoo, DeepEval, Ragas, and OpenAI Evals: offline vs online evals, LLM-as-judge, CI integration, and dataset management for agent testing.

A fair, sourced comparison of the three LLM observability platforms teams reach for once agents hit production: Langfuse's open-source tracing and prompt management, Braintrust's eval-first workflow for regressions, and Helicone's drop-in proxy for logging and cost control. Architecture, pricing model, self-hosting, and which to pick by workload.

Dan Luu's new agentic coding essay is not another vibe check. It is a useful reminder that coding agents only compound when the test loop, review loop, and task-selection loop are stronger than the code generator.

A Show HN project claims large agent-cost cuts by rendering bulky context as images. The useful lesson is not the trick itself. It is that compression needs evals, byte-safety rules, and per-request accounting.

OpenAI's June deprecations put Agent Builder, hosted Evals, and reusable prompts on a November 30 shutdown path. Here is the practical migration plan: Agents SDK, repo-owned prompts, and eval receipts.

The Bayer and Thoughtworks PRINCE case study is a useful reminder that reliable agentic AI comes from context routing, traces, evals, monitoring, and human review, not from a better prompt alone.

Hex's data-agent lab shows the practical eval pattern AI teams should copy: compare candidates against stable baselines, keep receipts, and judge changes by task behavior.

Open-source LLM engineering platform: tracing, evals, prompt management, and datasets. Self-hostable, OpenTelemetry-native, with 50+ framework integrations.