SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Async tool calls let AI agents keep working while long tasks run

0
·2 views

A DEV Community article argues that AI agents should use non-blocking tool calls rather than a blocking run() function. Under the proposed pattern, start() immediately returns a task handle, status() reports progress, and stop() cancels the job, while results come back later as events. The author explains that blocking calls cause latency to pile up, make user instructions stale, and risk gateway timeouts during slow tasks like test suites. The design is drawn from an essay by the DeepSeek Harness team, which describes it as full-duplex and notes it is implemented but not enabled by default. The author presents it as an architectural pattern to emulate rather than an official product feature.

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 ·

Open-source 'story box' built on ESP32 uses a kids' storytelling device for daily learning

A developer released Marelleo, an open-source story box built on an ESP32 microcontroller for his children, and it gained notable traction on Hacker News. The project's key design idea is to attach a daily training routine to content kids already enjoy, using a device already in the living room rather than new hardware. The article argues that long-term engagement, not hardware performance, is the main bottleneck in training and rehabilitation apps, which often lose users within the first week. The piece extends the same principle to vision therapy, citing Amblyotube, a Meta Quest app for amblyopia that delivers different visual feeds to each eye while users watch video content. The authors note that the app is a supplementary training tool and not a diagnostic or curative treatment.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

A 3x3 median filter cut OCR errors on a book page, but the same filter backfired on a half-size scan

A developer tested whether a 3x3 median blur could stay as a fixed step before OCR, sweeping kernel sizes on one page of H. G. Wells's The Time Machine. At full size, the 3x3 filter cut Tesseract's character errors from 11 to 2, while larger kernels became harmful. When the same page was scaled to half width, the same 3x3 filter raised errors from 4 to 28, and no kernel helped. The author suggests the effect depends on how the filter's window compares with letter stroke width, though this is a hypothesis from a single page and single run. The results indicate the right preprocessing depends on scan resolution rather than being a constant.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How pg_trgm fixed empty product search for misspelled Hebrew and English queries

The ramo.co.il tools catalog returned no results when customers mistyped a word, because its ILIKE matcher required an exact substring. Product titles mix Hebrew text with Latin brand names and model numbers, and shoppers often combine scripts or misspell each half independently, which made the problem common rather than rare. Simple synonym lists and Hebrew stemming did not help, since the errors were typos rather than inflections. The team adopted PostgreSQL's pg_trgm extension, which scores strings by shared three-character slices and so still gives partial credit for a single wrong letter. They store one concatenated search column per product and combine whole-string and word-level similarity scores, with a 0.3 threshold for queries of four or more characters. Very short queries stay on prefix matching, and a keyboard-layout mapping covers queries typed in the wrong language.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Piping a script through tail hides its failing exit code from AI coding agents

A solo iOS developer in Japan found that Claude Code reported his check script as passing when it had actually failed. The agent ran the script piped into tail, so it saw tail's exit code of 0 rather than the script's exit code of 1. Two known failures were involved the first time, and a new failure in a request-validation check appeared during an unattended session the second time. The article explains that a pipeline's exit status comes from its last command, and shows how to recover the first command's status in bash or zsh using PIPESTATUS or pipestatus. It also covers set -o pipefail, which works in both shells, and notes that GitHub Actions' default bash shell behaves differently from an explicitly set one.

0 comments Read more at DEV Community