87% of Custom Claude Code Agents Went Unused for 30 Days, Wasting Tokens

·3 views

A developer running an automated Claude Code setup discovered that seven of their eight custom-defined agents recorded zero invocations over a 30-day period from late May to August 2026. Only the code-reviewer agent was called even once, while specialists like architect, security-reviewer, and database-reviewer sat completely idle. The problem is structural: Claude Code injects all agent definitions into the system prompt on every request, meaning unused agents silently consume tokens and can degrade inference quality. Claude does not autonomously select agents based on their own descriptions — an agent only activates when explicitly invoked by a calling prompt or logic. The developer's takeaway is that agent utility should be measured through actual usage logs, not assumptions, and that unused definitions should be pruned regularly to reduce both token costs and management complexity.

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)

Developer builds browser-based image converter that keeps files off servers

A developer built ImagArt AI, a free image conversion tool that processes files entirely within the user's browser without uploading them to any remote server. The tool uses built-in browser APIs such as Canvas, toBlob(), and toDataURL() to decode, resize, and re-encode images to JPG, PNG, or WebP formats. It supports batch uploads and ZIP export, with all processing handled by the user's own device CPU. The motivation was to address privacy concerns with mainstream free converters, which typically require files to be sent to external servers. The developer noted that while the client-side approach eliminates server costs and privacy trade-offs, processing very large batches of high-resolution images can strain performance without careful use of Web Workers.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Filtering Out Unstable Audio Clips Fixes Hoarse TTS Style Synthesis

A text-to-speech system that builds emotional style vectors from a small set of representative audio clips was producing consistently hoarse output due to low-quality clips being included in the selection. Because only five clips are used per style, even one rough clip contributes 20% weight to the average style vector, pulling it toward abnormal acoustic regions. The developer identified two key audio quality factors — jitter (unstable vocal cord vibration) and octave jumps (sudden F0 pitch shifts) — and built a scoring function to measure both. Clips are now ranked by a combined stability score, with octave jumps weighted twice as heavily as jitter due to their greater perceptual impact. Replacing the naive first-five selection with this stability-based filter eliminated the hoarseness entirely across all synthesized output.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Probability Theory Unifies All Major Deep Generative Model Families

A technical explainer published on DEV Community argues that generative models are best understood through probability theory rather than as isolated architectures. The article frames generative modeling around three core problems: how to represent high-dimensional distributions, how to train a model distribution to match real data, and how to reason backward from observations to hidden variables. This unified lens, the author contends, connects autoregressive models, VAEs, GANs, flow-based models, and diffusion models despite their architectural differences. Unlike discriminative models that learn conditional mappings for prediction, generative models aim to approximate the full probability structure underlying the data. The piece also links generative modeling to classical inverse problems, where the goal is inferring likely causes from observed outputs.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Missing --cdp flag silently launches a second Chrome, mimicking a logged-out session

A developer debugging a scheduled automation job discovered that omitting the --cdp flag in agent-browser caused the tool to silently launch a fresh, unauthenticated Chrome instance instead of attaching to an existing logged-in one. The automation still ran successfully by all exit-code checks, but returned a logged-out page because it was operating in a brand-new browser profile with no stored session. This made the issue appear identical to a genuine session expiry, leading to misdiagnosis. The key diagnostic tell is the log line 'launched browser' — its presence confirms the tool is not attached to the user's actual Chrome profile. A reliable fix involves verifying the correct Chrome process is bound to the expected debugging port by inspecting its user data directory via the system process list.

0 comments Read more at DEV Community