Urgent.News

What's breaking now, across thousands of outlets.

Tech

The Agent That Polled 55 Times to Learn Nothing

Why a lead agent in a multi-agent pipeline kept sleeping, polling and ignoring its human, what it cost in tokens, and how a flatter setup fixed it.

The Agent That Polled 55 Times to Learn Nothing

In a multi-agent delivery pipeline, an agent named Lead experienced significant delays and inflated costs due to inefficient waiting practices. The Lead, which owned a ticket board and coordinated worker agents, relied on self-imposed waiting periods to check for task completion. This led to excessive sleep intervals, resulting in long periods of inactivity and missed communications.

The issue was identified through analysis of the logs, which revealed a pattern of prolonged sleeps followed by status checks on incomplete tasks. The Lead made 303 tool calls, with 55 of them being unnecessary sleeps, accounting for nearly 20% of the total. Each sleep interval lasted for more than 100 seconds, causing the Lead to become unresponsive and unable to process corrections or new information.

Once the inefficiency was discovered, the solution was straightforward: the Lead needed to wait for tasks to complete before checking their status. By reducing the number of unnecessary sleeps, the Lead could efficiently monitor the ticket board and respond to corrections promptly. This change resulted in a significant reduction in processing time, with the Lead spending only 3.66 hours in sleeps during a 19.3-hour period instead of the previous 3.66 hours.

Additionally, the excessive wakes up caused by the inefficient waiting led to a high token usage, accounting for over 25% of the total token consumption. If the Lead had simply waited for tasks to finish and checked their status 5 times, the token usage would have been around 1.2 million, rather than the observed 13.5 million.

In summary, the Lead's inefficient waiting practices resulted in excessive sleep intervals, missed communications, and inflated costs. By adjusting the waiting logic to check task completion promptly, the Lead was able to significantly improve its performance and reduce costs.

Written by urgent.news from HackerNoon's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at hackernoon.com →

More in Tech

Why Your 10 W Attenuator Failed at 5 W

A 10 W attenuator running at 5 W continuous should be the least interesting part on the bench. It still comes back. When it does, the report usually says "overpowered", and the test log says 5 W, and…

  • 10 W attenuator fails at 5 W due to real-world conditions
  • Operating temperature and VSWR affect power rating
  • Proper mounting and frequency considerations crucial

Azure Outage Triage, Browser MCP, Code Security, and Agent Tools

These five projects sit at different layers of the development stack, but they share a practical theme: giving developers better context before they change code.

  • Microsoft 365 Status and Azure Service Health separate incidents
  • Open Orange runs Cloudflare Workers on Azure without code changes
  • AI-assisted security validation gate identifies code vulnerabilities

AWS says data centers in space won't be viable for 10 or 20 years

Amazon Web Services CEO Matt Garman claims that data centers in space are "highly unlikely to be economically viable or practical in the next decade or two." Garman said as much on Alex Heath's…

  • AWS CEO Matt Garman doubts space-based data centers will be viable in 10-20 years.
  • Current space launch limitations and demand surge pose challenges for orbital data centers.
  • Deserts considered most probable site for data centers, but capacity shortfall expected.

More from Saturday 14 November →