12 Parallel Subagents, 444 Tool Calls: Running 3 Projects in One Day with Claude Code

444 tool calls. 3 sessions. 93 hours and 39 minutes of logged context. All in one calendar day.

One session published a three-part DEV.to series on trending AI GitHub projects. Another produced six browser-ready HTML market research reports on Korea’s dental advertising landscape. A third ran live PayPal transaction tests, pulled Japanese and Southeast Asian market data across 136 sources, and pushed a site redesign across 10 languages.

Three different projects. Three different domains. All running concurrently.

TL;DR Parallel subagent dispatch is the core lever. Spinning up 12 agents simultaneously collapses work that would take a full day into about 2 hours. But more agents also means more noise — without a validation layer, false positives scale with agent count.

”Post Some AI GitHub Project Analyses to DEV.to” — One Prompt, Three-Part Series

The first session started with a casual request:

analyze ~4 trending AI projects from GitHub and post articles to dev.to

Claude Code ran web searches to find trending projects as of April 2026: andrej-karpathy-skills (16K stars), hermes-agent, OpenClaw (295K stars), and opencode. Rather than writing four separate posts, it proposed restructuring them as a series: The 2026 AI GitHub Playbook.

Part 1 published immediately. Parts 2 and 3 went up as DEV.to drafts, queued for scheduling.

Published: https://dev.to/ji_ai/how-a-markdown-file-hit-16k-stars-skills-in-2026-36hi (DEV.to id=3542024, 2026-04-23 14:55 UTC)

There were exactly three user interactions in that entire flow: the initial prompt, “make it around 3 parts,” and “go ahead.” Web search, structure proposal, content generation, API publishing — all autonomous.

One behavior worth noting: an existing file claude-code-channels-vs-openclaw-en.md was already in the project. Claude Code detected it and approached OpenClaw from a different angle, linking internally rather than duplicating coverage. That judgment call wasn’t made by a human.

Midway through the same session, a separate issue surfaced. A quality gap between April 25 and April 26 posts on spoonai prompted patches to SKILL-2-publish.md and the self-critique.mjs validation loop. Two projects, one session. 191 tool calls — Bash 96, Agent 11.

12 Agents in Parallel: Market Research That Would Have Taken All Day Finished in 2h 26min

Session two was a dental advertising research project. The prompt:

research every company doing AI-assisted dental/medical advertising and write a report. use 10+ subagents

Twelve subagents launched in parallel, each assigned a non-overlapping domain:

• Korean AI medical advertising landscape (8 categories, 60 companies)
• Naver C-Rank / D.I.A.+ algorithm changes and their effect on medical content
• Active dental clinic blog case studies from real operators
• 5-year / 1-year / 90-day trend comparisons

Output: six HTML reports, all browser-ready.

• TREND-COMPARISON-REPORT.html
• AI-AGENCIES-DEEP-REPORT.html
• AI-AGENCIES-PRIMER.html
• AI-AGENCIES-EXAMPLES.html
• AI-AGENCIES-EVIDENCE.html
• AI-DENTAL-TELEGRAM.md

Evidence from each agent was tagged with a reliability tier. “Real name + quantitative data (5 stars),” “initials + rich metrics (4 stars).” One agency, Howon & Company, returned direct links to actual client case studies. Another, Inblog, surfaced a live dental clinic blog (heritagedental.inblog.io) as a working reference.

The structural key to making parallel dispatch work is domain separation. If multiple agents explore the same topic, results overlap and merging becomes expensive. Pre-partition the problem space so each agent has an exclusive slice, and you can concatenate outputs directly.

63 tool calls. Agent 35 times. 2 hours 26 minutes. The serial equivalent would have consumed most of a workday.

Telegram Message → DB Query → Live PayPal Test → Market Data → Design Push

Session three started as a Telegram message:

has anyone visited or paid for the saju project?

The Telegram plugin connects directly into Claude Code’s session context. A DB query ran immediately: 30 cumulative payments (₩171,000), no new payments since March, 87 sessions still arriving through April. Toss payment confirmed working. LS Pay dead — declined the fortune-telling category. PayPal live keys configured but never tested with a real transaction.

The next message:

spin up one agent each for Japan and Southeast Asia — figure out how to make money there

Four agents ran in parallel in the background:

• JP fortune market data → jp-market-data.md
• SEA fortune market data → sea-market-data.md (136 inline sources)
• Viral fortune video pattern decode → viral-formula.md
• Top-converting fortune site references

While the agents ran, the PayPal live endpoint got a direct test: a real $1.99 order created in the DB, approval URL issued, results written to scripts/paypal-live-test.sh.

One agent result flagged a real bug: common.json:3 had 運命研究所 while countries.ts:142 still read FortuneLab. Brand inconsistency across two files. An actual fix was needed.

Another agent raised a false positive: “Thai users can see the ₩ symbol.” Reading the code directly, ₩ rendering lives inside the toss namespace — Korean checkout only. Thai users route to a PayPal-hosted page and never encounter ₩. More agents means more noise. Without a validation layer, scaling agent count scales false positives at the same rate.

After market data came back:

update the design

i18n message files across 10 languages, page.tsx, paywall/page.tsx, and globals.css — all updated. 190 tool calls, Bash 92, Edit 25.

Tool Usage Across 3 Sessions

Session	Duration	Tool Calls	Top Tools
Session 1 (DEV.to + spoonai)	75h 58min	191	Bash 96, Agent 11
Session 2 (Dental ad research)	2h 26min	63	Agent 35, Bash 9
Session 3 (Saju global)	15h 15min	190	Bash 92, Edit 25
Total	93h 39min	444	Bash 197, Agent 53

44% of all tool calls were Bash. 12% were Agent. Fifty-three agent dispatches drove the parallel research that defined these sessions.

Tool Choice Is Strategy

Agents are most efficient in research-heavy work. Wide search surface, parallel collection, and even if individual agents produce errors, the aggregate result usually holds.

For code changes, direct Edit outperformed agents every time. Twenty-one i18n files got updated in 25 Edit calls — no agent coordination needed.

More agents also means more false positives. The Thai ₩ symbol flag was concrete: an agent raised an issue that a 30-second code read dismissed. Without a validation layer over agent output, noise grows with scale.

The pattern across these three sessions: use agents when the search space is large and results can be validated by inspection. Use direct tools when the action space is small and precision matters.

Domain separation determines output quality. Tool choice determines throughput.

---

More projects and build logs at jidonglab.com