89% of AI agent teams use observability, but fewer than half run evals

·1 views

A LangChain survey of 1,340 AI practitioners, published in May 2026, found that 89% had implemented observability for their agents, yet only 52.4% ran offline evaluations and 37.3% ran online evaluations. Fewer than one in three teams ran both types of evaluation. The core problem is that observability tools — which track latency, token usage, and span status — cannot detect whether an agent's output was actually correct. A confidently wrong action, such as refunding the wrong order, produces identical telemetry to a correct one, with all spans showing green and no errors logged. Despite this gap, the same survey identified output quality as the leading barrier to deploying agents in production.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Publora Lists on Claude Connector Directory After Month-Long Build and Portal Glitches

The team behind Publora successfully submitted their product to Anthropic's official Claude connectors directory, with the submission form itself taking only about fifteen minutes to complete. However, the process was preceded by roughly six weeks of preparation, including reviewing thirteen pages of documentation, testing the server against fifteen requirements, and manually running eighteen tools in production. After hitting publish, the portal immediately showed a 'live' status, but the public listing page returned a 'connector does not exist' error for several hours before resolving on its own. The team also encountered several other discrepancies, including the portal incorrectly flagging the server as requiring no authentication, four mismatches between the official documentation and the actual submission form, and all eighteen tools being flagged as missing titles despite having them. The author shared these issues publicly to help other developers avoid the same time-consuming pitfalls when submitting their own connectors.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Polars 2.0 parses 1M dates 25x faster than lru_cache in Mac mini benchmark

A hands-on benchmark run on a Mac mini M4 Pro with Python 3.14 compared three approaches to parsing one million date strings: a plain script, an lru_cache trick, and Polars 2.0.0-rc.1. The cache method delivered roughly 3x speedup over the baseline but only when input dates repeated, and was actually 6% slower than doing nothing on fully unique inputs. Polars 2.0 completed the same task in 0.13 seconds end-to-end — about 25 times faster than the cached version — regardless of whether dates repeated. A separate re-run of the uv package installer at version 0.12.11 showed it remains around 6-7x faster than pip, with only a 4% improvement over the previous 0.11.6 version. The author notes Polars 2.0 is still a release candidate and recommends rechecking results once the final version ships.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

hi3d-cli Lets Developers Generate 3D Models From Terminal Using AI API

A new open-source command-line tool called hi3d-cli enables developers to generate textured 3D models directly from the terminal without manual downloads from web dashboards. The zero-dependency tool connects to the Hi3D API and supports multiple output formats including GLB, OBJ, STL, FBX, and USDZ, covering use cases from web apps to 3D printing. Setup takes under a minute via npx, with an interactive wizard storing API credentials locally. The tool also integrates with Claude Code as an agent skill, allowing automated 3D asset generation within existing development workflows. The package is available on both GitHub and NPM under the handle hi3d-cli.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Free-List, Slab, and Buddy Allocators Solve Memory Management Challenges

A developer tutorial series exploring OS memory allocation explains how page-based memory management introduces new challenges around efficiency and overhead. Modern 64-bit CPUs read data in 8-byte chunks, making memory alignment critical for performance, though page-aligned memory from the OS helps address this. A key problem arises with small allocations, where metadata overhead can exceed 600% of the actual memory requested. The Linux kernel addresses this with Slab Allocators, which use a single control structure to manage hundreds of fixed-size slots. The article introduces three main allocator strategies — Slab, Buddy, and Free-List — each suited to different use cases, and demonstrates a basic Free-List implementation using linked lists and pointer arithmetic in C.

0 comments Read more at DEV Community