Projects About

4 Parallel Agents, 102 Tool Calls: Automating Naver Algorithm Research with Claude Code

Three dental ad clients asked the same question in the same week: “Naver feels different lately — did the algorithm change?” The ADVoost rollout had shifted how results behaved, and the feedback kept coming. “It seems like it changed” isn’t a useful answer. 6 sessions and 102 tool calls later, I had analyzed 17 official Naver announcements and produced 3 HTML reports.

TL;DR: Used the research skill to dispatch 4 parallel agents → Codex cross-verification to isolate unconfirmed claims → iterative HTML report generation. Bonus: caught the Naver AI Briefing ad beta launch in real time, one day after it went live on 2026-05-07.

4 Opus Agents Simultaneously: 4 Research Angles in 21 Minutes

Single-prompt research on Naver algorithm changes doesn’t work. Official announcements, industry community observations, ADVoost technical changes, dental ad practice implications — these four angles pull from different sources with different reliability levels. Mix them in one prompt and the model blends verified facts with speculation.

The research skill’s solution: dispatch 4 agents in parallel from a single message.

Agent 1: Official announcement extraction & Search Advisor docs
Agent 2: Organic search & Place ranking changes (community + industry observation)
Agent 3: ADVoost matching logic analysis
Agent 4: Dental/medical ad practical implications

Each agent prompt required: “source URL mandatory, flag overlap with sibling agents.” Once 4 independent output files landed, a separate synthesis session extracted agreements, contradictions, and unverified claims. Work that would’ve taken 80 minutes sequentially finished in 21.

Session 1 tool calls: Agent(8), Read(5), Bash(4), TodoWrite(3), Write(2).

The Unverified Announcement Codex Caught — Evidence Grading Introduced

The problem surfaced immediately when combining all 4 agent outputs. The first draft included a “2026-05 Place Ad policy announcement” stated as fact — but it appeared in none of the 15 entries in naver_ads_notice_extracts.json. In medical advertising, stating “algorithm changes” without official confirmation is a compliance liability.

Session 2 created a standalone claude_synthesis_review.md and enforced five evidence tiers:

CONFIRMED      → explicitly stated in Naver Ads notices
               (broad match, ADVoost, medical content rules, Place ad test)
INFERRED       → reasonably deducible from official help docs
OBSERVED       → community/practitioner reports, no official notice
EXTRAPOLATED   → logical reasoning from confirmed data
UNVERIFIED     → no source or unconfirmed claim

Organic search and general Place ranking are undisclosed by Naver policy — they couldn’t be elevated above OBSERVED. Without this explicit grading step, unconfirmed claims would’ve made it into the client report unchanged.

Session 2 tool calls: Read(4), Bash(4), Write(2), Grep(1). Elapsed: 5 minutes.

API Overload and Why File-Based State Management Actually Matters

Session 3 hit API Error: Overloaded attempting the first HTML report generation. Opus 4.7 with 8 parallel agents was too many concurrent requests.

No work was lost. research-minutes.md and claude_naver_research_report.md were already written to disk. Session 4 picked up immediately and moved straight to report generation. This is the practical payoff of file-based state management: when agents write intermediate results to files, a broken session doesn’t lose context.

Session 4 produced naver_algorithm_ad_agency_prediction_report_2026-05-08.html — 40.9KB, 429 lines. Structure:

§1 Key conclusions + medical ad safety box (pinned to top)
§2 Evidence tier legend (5-level visualization)
§3 Objective data — metric cards from 17 announcements
§4 Organic search & Place observations
§5 Ad agency impact predictions
§7 Dongbaek Place Ad pilot analysis

Session 4 tool calls: Bash(15), Read(6), Grep(4), Glob(1), Write(1). Bash ran 15 times because each Write was followed by script-based validation — tag balance, DOCTYPE presence, viewport meta, section existence.

The client then requested “last 6 months only, essentials.” Session 5 produced a third version. Three HTML reports across three sessions wasn’t scope creep — requirements evolved: “overall changes” → “agency impact predictions” → “last 6 months compressed.”

The Daily Agent That Caught Yesterday’s AI Briefing Beta

Session 6 produced an unexpected result. Running the daily research agent defined in medical_dental_ads_daily_goal.md, one of the official source scans fired a hit.

2026-05-07 — the day before — Naver launched an ‘AI Briefing’ ad beta.

The key was explicit target URLs in the agent prompt:

- Naver Ads Notice:         https://ads.naver.com/notice?categoryId=147
- Naver Ads Help:           https://ads.naver.com/help
- Naver Search Advisor:     https://searchadvisor.naver.com/
- Korea Dental Ad Review:   https://www.dentalad.or.kr

Without this list, generic WebSearch prioritizes blog posts. Specifying URLs directly changes the source reliability floor.

WebSearch(9) + WebFetch(6) confirmed:

  • Phase 1: Shopping search ads/ADVoost integration, shown below AI-generated answer summaries
  • Healthcare vertical AI: explicitly on the in-year roadmap
  • Medical keyword coverage: [UNVERIFIED]

Without the daily agent, this would’ve surfaced days later. Instead it went directly into that day’s client briefing.

Full Stats

ToolCallsPrimary use
Bash29File validation, state updates
Read28Source file reads
Write9Output file generation
WebSearch9Official source discovery
Agent8Parallel research dispatch
WebFetch6Announcement page crawling
Grep5Keyword verification
TodoWrite3Step tracking
Total102

8 files created, 0 files modified. Each iteration produced a new file rather than overwriting — overwriting intermediate results removes the comparison baseline for Codex cross-verification.

Parallel agents work well for breadth problems. 4 simultaneous agents outperform 4 sequential runs on both speed and coverage balance. Evidence grading has to be an explicit step — without a dedicated classification pass, verified facts and inferences merge. Moving the grading to a separate session and running Codex cross-verification measurably improved the final report quality.


More projects and build logs at jidonglab.com

Comments 0

0 / 1000