SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Saying 'Hi' to an Internal AI Agent Consumed 177,000 Tokens Due to Tool Definitions

0
·2 views

Bold.org's internal company agent, built on OpenCode, saw token usage climb faster than its session count could explain. A fresh conversation containing only the word "hi" was recorded at 177,444 input tokens, which ruled out long conversations as the main cause. Analysis with a token-tracking tool pointed to the skill catalog and tool definitions, which the tool's inventory did not fully capture. Investigators traced the bulk of the overhead to full MCP tool definitions, particularly a Meta Ads integration with 97 tools whose descriptions and schemas were sent with every request. The case shows how integration documentation loaded into an agent's context can dominate costs even when unused.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

NVIDIA releases Nemotron 3.5 Transcribe for local, single-GPU speech transcription in 40 languages

NVIDIA's Nemotron 3.5 Transcribe can run on a single GPU to transcribe hour-long recordings in 40 languages without per-minute fees or sending audio to outside servers. The model tops the Hugging Face Open ASR Leaderboard among open-weight models and is licensed under OpenMDW-1.1, which permits commercial use. NVIDIA published BF16, FP8, and NVFP4 weights on its October 28, 2026 release date. The guide covers hardware requirements, three ways to run the model (vLLM, SGLang-Omni, and Transformers), and common setup mistakes. Hardware compatibility depends on the GPU, and 24GB cards and Apple Silicon are not supported.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Template Literals Favored Over + and .join() for JavaScript String Concatenation

A DEV Community article compares three ways to join strings in JavaScript: template literals, the + operator, and Array.join(). It recommends template literals as the default choice for most cases because they are readable, avoid quote escaping, and support multi-line strings. The + operator is described as suitable for simple single-line combinations, while .join() suits large loops or building strings from arrays. The author notes that older browsers made + slow in loops, but modern engines like V8 optimize it, so template literals and + are generally fast enough for everyday use. The piece also promotes a client-side utility for converting between concatenation styles.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

New DNSSEC guidance requires DS automation to verify CDS and CDNSKEY across all nameservers

RFC 10026, published in July 2026 as BCP 246, sets out operational recommendations for automating DNSSEC Delegation Signer updates. It requires that CDS and CDNSKEY records be consistent with each other and across every authoritative nameserver for a delegation, and that the resulting DS set would still permit validation. If either check fails, the parent must cancel the change. The author has published a standard-library Python script implementing these checks, which was run on 29 October 2026 against four zones publishing CDS records. The script builds on RFC 7344, RFC 9975 and RFC 8078, which define acceptance rules, consistency tests and the delete signal.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Running NVIDIA's Nemotron 3.5 Transcribe Locally on a Single GPU

A DEV Community guide explains how to run NVIDIA's open-weight Nemotron 3.5 Transcribe model on one GPU to transcribe hour-long audio files in 40 languages. The model, released October 28, 2026 under the commercially usable OpenMDW-1.1 license, currently tops the Hugging Face Open ASR leaderboard. The article argues local deployment avoids per-minute API fees, which it says can total about $1,200 for 10,000 hours, and keeps sensitive recordings within a user's own network. It covers hardware compatibility across NVIDIA GPUs, the BF16, FP8, and NVFP4 checkpoints, and three runtimes: vLLM, SGLang-Omni, and Transformers. The guide also notes that Apple Silicon and 24 GB cards like the RTX 4090 are unsupported.

0 comments Read more at DEV Community