← /editions
#144

Your Background Agents Are Talking. A Lot.

// A gist reports Claude Code regularly sends each background agent's transcript back to its model, and a teardown measures what Claude Code's multi-agent code review costs and finds. Multi-agent setups carry coordination costs teams should measure before scaling.

$ today_in_10_seconds

A developer's gist reports that Claude Code regularly sends each background agent's transcript back to its model. A separate teardown measures what Claude Code's multi-agent code review costs and what it catches. Running many agents at once is now common, and today's question is what that orchestration actually costs.

$ lead_story

Claude Code regularly sends each background agent's transcript to its model

A gist on Hacker News describes Claude Code repeatedly passing each background agent's transcript to the model rather than letting agents run as fire-and-forget workers. If it works as described, every parallel agent you start adds recurring context, so token use grows with both the number of agents and the length of each transcript. That's a sensible design for keeping an orchestrator aware of what its workers are doing, but it's a cost most people won't see until the bill arrives. If you're building on agent harnesses, test how your tools handle subagent context before you scale to ten workers.

// why_it_matters: Starting another agent costs more than one agent's worth of tokens if its whole history keeps getting sent back.

read more →

$ also_worth_attention

  • What Claude Code's multi-agent code review costs and finds

    A hands-on teardown of Claude Code's multi-agent review setup, covering both the spend and the bugs it actually surfaces. This is the comparison that matters for any team deciding whether parallel reviewer agents beat a single careful pass. Check whether the extra findings are real defects or just more noise.

    link →

$ bigger_picture

The trend behind the headlines

Both stories make the same point: multi-agent setups don't add up linearly. Coordination has costs, like transcripts sent back to the model, overlapping reviewers, and orchestrators keeping track of every worker. The market is moving from 'can agents do this?' to 'is it worth paying for N agents to do it?' Teams that measure cost per useful finding, not just total findings, will pull ahead of teams that add more agents and hope for the best.

$ tool_of_the_day

ccusage

An open-source CLI that reads Claude Code's local usage logs and breaks down token use and estimated cost by day, session, and model. After today's stories, it's the quickest way to see what your background agents and review runs actually consume.

explore ccusage →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime