QWE AI Academy The status line that quietly eats your quota
Here’s the bit most Agent View walkthroughs skip: Claude Code regularly sends each background agent’s transcript to its model so a short status line can say what that agent is doing – often just three to five words. Those lines look free. Community OpenTelemetry breakdowns from late October 2026 put related agent_summary traffic around one-fifth of total token spend in heavy multi-agent runs (about 16.4% in one published Flamerix set; independent checks in GitHub issue #86879 landed roughly 18-26%).
That discovery is what’s blowing up right now. The UI still feels magical. The meter tells a colder story.
What actually gets sent (two different paths)
Don’t mash every “summary” into one bucket. Claude Code has at least two related behaviors, and only one is fully spelled out in the product docs.
| Path | What you see | When the model is called | What docs say |
|---|---|---|---|
| Agent View row text | One-line status per background session | Fresh write at end of turn; mid-turn rewrite every few minutes on long turns | Haiku-class request, billed like the session; 15s ticks reuse local output with no model call |
| Agent progress / AgentSummary | Tiny “what is this subagent doing” chip in panels | Periodic timer (community reverse-engineering points at ~30s defaults, server-tunable) | Not called out the same way in cost docs; community reports heavy input with tiny present-tense output |
Row summaries exist so you can skim without opening the full transcript – working rows describe the current action; blocked rows surface the question waiting on you (Agent View docs). Progress-summary traffic is the path community telemetry keeps tripping over: fork-ish requests that re-read a lot of agent context to emit a fragment like “Reading runAgent.ts.” Those answers never land in your main chat. They decorate a status surface and vanish.
Hands-on: watch the pipeline without guessing
You don’t need a conspiracy board. A ten-minute lab is enough.
1. Spin two independent background sessions
claude agents
# type a concrete task, press Enter
# example: "list failing tests in this repo and open a short report"
Or from a shell without opening the view first:
claude --bg "map error handling in packages/api and note gaps"
claude --bg "summarize recent git history for auth/"Each Enter in Agent View starts a new session, not a follow-up. Official limitations are blunt about the bill: ten parallel background sessions burn usage roughly ten times as fast as one.
2. Separate free ticks from billed rewrites
Leave Agent View open and watch one Working row. Text can refresh about every 15 seconds from the session’s own recent output – no model call. When a turn ends, or a long turn drags, the row gets a model-written line. That second kind is the short Haiku-class request Anthropic documents.
Pro tip: On third-party providers or gateways with no Haiku-class model, row rewrites fall back to the session’s main model. Set
ANTHROPIC_DEFAULT_HAIKU_MODELor one-line UI text can price like Sonnet or Opus.
3. Measure what /usage can’t fully show
/usage (or /cost) inside an attached session is only a partial picture – local list-rate estimates unless your org injects contracted pricing. Turns out some summary-side requests never show up cleanly in the on-disk JSONL under ~/.claude/projects/. Grep the pretty session file alone and you’ll under-count. The 15-20% band showed up by grouping API events on query_source (for example agent_summary) in OpenTelemetry-style exports – not from transcript files. Session logs stay useful for “what did the agent do”; data-usage docs still describe plaintext local caching with default cleanup around 30 days. Weak instrument for invisible side queries.
4. Decide what “good enough” status is worth
- Live steering across many independent jobs? Keep Agent View. Space peeks, arrows move rows, Enter/→ attaches, ← detaches. Use state icons and PR labels; ignore three-word chips you never act on.
- Parallel work without the dashboard tax? One session with targeted subagents that return a single summary, or fewer longer-lived background sessions – not a swarm of short ones.
- No background orchestration UI?
disableAgentVieworCLAUDE_CODE_DISABLE_AGENT_VIEWkills the surface (and its documented row-summary path), not every other agent feature.
Common pitfalls that look like “Claude got expensive overnight”
Counting only main-chat turns misses the stack. Parallel agents multiply full agent loops; status summarization sits on top. You can do everything “right” in the foreground and still watch a five-hour plan bar melt because five detached workers keep waking a summarizer.
Hunting for a master mute wastes time. As of the late-October 2026 write-ups, there is no blessed “disable agent progress summaries” switch. Undocumented CLAUDE_CODE_AGENT_PROGRESS_SUMMARIES is described as enable-only. Env vars that sound like they mute non-essential calls either weren’t in the binaries people inspected or only help by shutting background agents down entirely. Community Bun preload hooks exist; they’re fail-open, version-fragile, and a compliance call – not a first-class product control. (LavX’s reconstruction of the AgentSummary path and Flamerix’s fail-open warning are the clearest public notes – see the October 2026 LavX report.)
Is a three-word status worth re-reading a fat slice of transcript every half minute? Depends how many agents you leave running while you make coffee.
Performance and results: what the numbers actually imply
Enterprise Claude Code sits around $13 per developer per active day and $150-250 per month, with 90% under $30/active day – figures from Anthropic’s costs page as of October 2026 reads. Stack community summary findings on top: teens of percent of API spend on multi-agent days, on a path optimized for UI chrome. Fix the shape of the work. Typing faster does nothing here.
What actually moved the needle:
- Cap concurrency. Two serious background sessions beat eight idle ones painting status chrome.
- Lean on Agent View’s free 15-second local refresh plus end-of-turn Haiku row text – not constant micro-updates nobody reads.
- Keep tasks short enough that mid-turn “every few minutes” rewrites don’t stack all afternoon.
- On API billing, route UI chrome to real Haiku. Haiku 5.5 list rates (late 2026 public tables) sit near $0.10/$0.50 per MTok under the 100K prompt tier, higher above it – far under Sonnet/Opus.
- Re-check telemetry after upgrades. Summarizer intervals and “short mode” (~2k token notes instead of the full transcript) have moved via server flags; v2.1.310’s cheaper short path was reported default-off, so client users can’t opt into it yet.
When you should not lean on this
Tight Pro five-hour window and the jobs aren’t independent? Skip the fleet. Opened Agent View just to feel productive? Each row is a full session, not a thread – close it. Gateway bills summary traffic at the main model with no Haiku route? You’re paying frontier rates for chrome.
Depth over dashboard density: one interactive session, or one subagent that returns a single result.
FAQ
Does Claude Code send the full transcript to Anthropic just for status text?
No separate “upload my disk” pipe for Agent View rows. Docs describe short Haiku-class requests built from session activity, billed under normal data-usage terms. Community-described progress summaries reuse agent context to mint UI-only phrases. Different from the explicit post-session transcript share prompt.
How do I see the cost without OpenTelemetry?
Partial picture only. Attach, run /usage, watch plan bars, compare a one-agent day to a five-agent day. Without query_source exports, treat multi-agent days as structurally more expensive and design concurrency down – don’t wait for a perfect line item.
Is turning off Agent View the same as stopping transcript summaries?
No – and this is where people get false comfort. Killing the dashboard removes that surface and its documented row-summary path. Background subagents inside a normal session can still drive other progress UI. Default-on interactive progress summaries (community timeline: roughly since 2.1.232 / Aug 2026) still lack a documented off switch. The clean product lever today is fewer concurrent agents and shorter unattended runs, not a single master mute the docs guarantee.
Next action: open claude agents, dispatch exactly two real tasks, leave the view up for one full turn cycle, run /usage on attach, and write down whether the status chrome changed any decision you made. If it didn’t, cut concurrency before you chase exotic hooks.