Agent Token Efficiency

Agent Token Efficiency Skill

Use this skill when the user wants to save tokens, reduce context bloat, configure coding agents, propagate skills/tools/profiles/MCP servers, analyze usage, or design provider/model rotation for continuous agentic coding.

This skill codifies a layered strategy:

1. Premium models make architecture and final tradeoff decisions.
2. Cheap subagents gather evidence, summarize logs, research docs, and map impact.
3. MCP servers are enabled sparingly because every exposed tool adds context overhead.
4. Skills and profiles are preferred over giant global instructions.
5. Usage metrics drive the next optimization pass.
6. Provider rotation uses measured quotas, not vibes wearing a lab coat.

Included CLI

Run the bundled Typer CLI:

python scripts/agent_efficiency.py --help

Recommended install from this skill directory:

python -m pip install -e .
aeff --help

Core workflows

1. Optimize an unsupported agent

Given the path to an unknown agent CLI or repository, inspect its capabilities, discover docs hints, map supported strategies from Claude/Codex/Kimi/Hermes, and produce/apply a token-efficiency optimization plan.

aeff optimize-unknown /path/to/agent-or-executable --docs-url https://example.com/docs --out ./reports

Use --apply only after reviewing the report:

aeff optimize-unknown /path/to/agent-or-executable --apply

The workflow looks for:

• config files
• model/provider settings
• MCP or tool registration
• custom agents/subagents
• skills/prompts/profiles
• compaction/summarization settings
• approval/sandbox controls
• usage accounting hooks

It then applies equivalent strategies where safe:

• add scout/research/log-compression workers
• add profiles for deep/cheap modes
• reduce always-loaded instructions
• add docs/code-search MCP only when useful
• add usage-reporting hooks
• disable unsafe auto-approval by default

2. Propagate a new tool, skill, profile, MCP server, or general update

Apply one update across the supported CLI fleet.

aeff propagate-update --kind mcp --name context7 --source assets/mcp/context7.json --target all

Supported --kind values:

• mcp
• skill
• profile
• agent
• instruction
• config

Dry-run is default. Add --apply to mutate files or run agent CLI commands.

3. Analyze token usage and propose improvements

Read local usage metrics, inspect configured agents/providers, compare provider free/throttled capacity, and calculate a rotation plan.

aeff analyze-usage --providers assets/configs/providers.yml --desired-rpm 4 --desired-daily-requests 500

With web refresh of provider notes:

aeff analyze-usage --allow-web --providers assets/configs/providers.yml

This workflow should produce:

• detected usage hotspots
• agents/models likely overused
• MCP/tool bloat warnings
• suggested model/profile changes
• free/throttled provider rotation plan
• explicit quota assumptions

Safety rules

• Default to dry-run.
• Do not commit secrets.
• Do not enable YOLO/auto-approval globally.
• Do not install large MCP suites by default.
• Prefer official docs for provider/rate-limit facts.
• If a rate limit cannot be verified, mark it unknown and require manual input.
• Never recommend violating provider ToS by creating throwaway accounts or evading limits.

Files

• scripts/agent_efficiency.py — Typer CLI implementation.
• assets/configs/providers.yml — provider quota and rotation assumptions.
• assets/mcp/*.json — bundled MCP server definitions.
• assets/agents/* — reusable scout/reviewer/research agents.
• assets/templates/* — AGENTS.md, CLAUDE.md, and reports.
• references/*.md — implementation playbooks.

Invocation guidance for agents

When this skill is active:

1. Use aef doctor or python scripts/agent_efficiency.py doctor before modifying a machine.
2. Use dry-run reports first.
3. Apply changes only after confirming file paths and generated diffs.
4. After changes, run smoke checks.
5. Store reports under ~/.agent-token-efficiency/reports or the requested output directory.