Agent Blast Radius

Agent Blast Radius

Periodically (or on demand) inventory the actual footprint of an autonomous agent: not what it was allowed to do, but what it has done. Designed for long-running Claude Code agents, scheduled tasks, and pipeline workers where the cumulative footprint silently expands and nobody notices.

Addresses Nate's "no executive visibility, no IT involvement" failure mode: by day 90, the agent has touched far more of the system than anyone realized.

When to Use

• Weekly review of any persistent agent (the agent platform, scheduled cron tasks, mission-control workers)
• Before merging a long-lived agent branch back to main
• After any incident where you need to answer "what did the agent do in the last 24 hours?"
• As a scheduled task itself — run the blast radius report on yourself

Inputs

• Project root (defaults to cwd)
• Time window (defaults to last 7 days)
• Optional: agent identifier (for filtering git authorship, log streams)

Phases

Phase 1 — Filesystem footprint

git log --since="<window>" --author="<agent>" --name-status --pretty=format:"%H|%ai|%s"

Capture: commits authored, files modified (added/modified/deleted), lines changed. Group by directory to show concentration.

Phase 2 — External call footprint

Check for evidence of external API calls in the time window:

• Grep ~/.claude/projects/*/logs/ (or equivalent) for tool-use entries
• Inspect any rate-limit/cost-tracking SQLite DBs the project maintains (e.g. store/the agent platform.db token_usage table)
• Read scheduled-task tables for tasks the agent created during the window

Phase 3 — Process / service footprint

• Active pm2 / systemd / cron entries created or modified by the agent
• New ports opened, new background processes spawned (where logged)
• Docker containers / stacks created via the container manager API (if relevant)

Phase 4 — Data footprint

• Rows inserted/updated in tracked SQLite/Postgres tables during the window
• Files written outside the project root (config files, logs, vault notes)
• Git branches/PRs created on remote orgs

Phase 5 — Produce the report

AGENT BLAST RADIUS REPORT
=========================
Agent: <id>
Window: <start> → <end>

FILESYSTEM
- Commits: N | Files touched: N | Net lines: +X / -Y
- Hot directories: <list with counts>
- Files outside project root: <list>

EXTERNAL CALLS
- API calls: N (Anthropic: N, OpenAI: N, ...)
- Estimated cost: $X.XX
- Rate-limit headroom remaining: X%

PROCESSES & SERVICES
- New scheduled tasks: N
- New pm2/systemd entries: N
- New containers: N

DATA WRITES
- DB rows touched: N (table: count, ...)
- Remote branches created: N (PRs: N)

EXPANSION TREND
- Footprint vs prior window: +X% / -Y%
- New surface area introduced: <summary>

ESCALATIONS
- <items requiring human review, e.g. files outside expected scope>

Verification

• Diff two consecutive runs; expansion trend should match recent activity
• Cross-check filesystem counts against git log --shortstat for the same window
• For Claude Code agents, verify token costs match the cost-tracking DB

Source Attribution

Concept extracted from Nate's Newsletter, 2026-04-05: "Executive Briefing: OpenClaw Deployments Are Spreading Through Your Org" — specifically the "no executive visibility, no IT involvement" pattern and the day-90 footprint expansion problem.