The hidden computational cost of autonomy
Autonomous AI agents promise to take high-level goals and solve them without human intervention. An editor might ask an agent to monitor a meeting, pull transcripts, and write a digest. That task is more complex than asking for a single headline. But this open-ended freedom requires massive computational power.
Every interaction with a model relies on tokens, the basic units of text and code that AI systems process and bill for. A single agent uses roughly four times more tokens than a standard chat. Multi-agent systems are even hungrier, consuming 15 times more. When multiple bots talk to each other to plan and execute tasks, token usage climbs quickly.
Much of this extra spending happens behind the scenes. A study of multi-agent software work revealed that the review-and-verify phase consumed nearly 60% of total tokens, primarily due to context passing between agents. Rather than producing usable journalism, the models spend most of their budget checking each other’s work and repeating background information across the system.
Structured workflows beat agentic chaos
The gap between autonomous agents and structured workflows was highlighted in a recent experiment summarizing a multi-day conference. In the first test, called Build 1, the task was handed to the Goal feature in Codex using a one-line objective. The system received no specified method, no list of target videos, and no definition of done.
Codex ran for just under 15 minutes. It used the iOS client to pull captions and wrote a throwaway parser in temporary storage. In the end it produced a nine-theme synthesis from 39 transcripts out of 41 videos found. Without strict guidelines, the autonomous agent hallucinated its source material. It bundled videos from a completely different conference year into its summary.
“You’re right. I over-included by treating the continuous channel block before Welcome to NAMS25 as all NAMS26. I’m correcting the source set to the first 23 videos only and will restate the recap against that narrower set.” - Codex, autonomous agent
By contrast, the experimenters ran Build 2 using a strictly scripted workflow. This process used standard tools like yt-dlp (an open-source video downloader) and youtube-transcript-api to fetch JSON transcripts-a common data format for text. It then assigned the lightweight Claude Haiku model to create structured per-session summaries. Those clean files went to Claude Sonnet to synthesize across sessions. The deterministic pipeline eliminated the hallucination error entirely.
This is the heart of the trade-off. Build 1 felt like magic: one sentence in, a finished report out. But the magic hid the risk of silent mistakes. Build 2 looked like engineering: a fixed sequence of tools and models, each doing one defined job. For a newsroom, that matters. A mislabeled video count in an automated summary could mean publishing the wrong source attribution, cheaply and at scale.
Engineering context instead of tweaking wording
This transition has pushed engineering teams to look beyond traditional prompt engineering. In 2025 guidance on model strategies, Anthropic described context engineering as the natural progression after prompt writing. The company defines this as the careful selection and maintenance of optimal tokens during inference-the process of a model generating a response. This ensures the model sees only the information needed for a specific step.
At the same time, harness optimization-improving the scaffolding, environment, and code tools surrounding the model-delivers measurable accuracy gains. In practice, that can mean a newsroom’s content team wires its internal archive directly into a summarization workflow instead of pasting copied source text each time. It can mean standardizing how meeting minutes are transcribed and filtered before a model ever reads them. The work happens in the plumbing, not in the phrasing of the request.
Research from 2026 on harness optimization showed that these tweaks helped raise the Terminal-Bench 2 pass@1 score from 69.7% to 77.0%. Anthropic’s 2024 guidance for practitioners states that agent-based systems typically cost 3 to 10 times more than workflows for deterministic tasks. Deterministic tasks are processes that produce the same output for a given input. As publishers face tight budgets and rising technical demand, newsroom leaders must decide between the apparent ease of autonomous goals and the proven reliability of structured workflows.
Written by Maciej Żemojcin using the Tribune Desk AI platform. Every claim in this article was fact-checked against its sources, and an editor read, edited and approved it before publication.
