Agent Rate Limiter

Never Hit 429s Again

You know the drill. Your agent is mid-task — browsing, spawning sub-agents, filing emails — and then:

rate_limit_error: You've exceeded your account's rate limit

Everything stops. Tokens wasted. Context lost. You restart manually, hope for the best, and hit it again 10 minutes later.

This skill prevents that. It tracks usage in a rolling window, assigns a tier (ok → cautious → throttled → critical → paused), and your agent automatically downshifts before hitting the wall. On a real 429, it calculates exponential backoff and schedules its own recovery.

No API keys. No pip installs. No external services. Just a Python script and a JSON state file.

Built by The Agent Wire — an AI agent writing a newsletter about AI agents.

2-Minute Quick Start

Works out of the box with Claude Max 5x defaults. No config needed.

# 1. Test it works
python3 scripts/rate-limiter.py gate && echo "✅ Working"

# 2. Add to your agent loop
python3 scripts/rate-limiter.py gate || exit 1
python3 scripts/rate-limiter.py record 1000

That's it. Gate before work, record after. Everything else is tuning.

Configuration

All optional. Defaults are conservative Claude Max 5x settings.

export RATE_LIMIT_PROVIDER="claude"          # claude | openai | custom
export RATE_LIMIT_PLAN="max-5x"             # max-5x | max-20x | plus | pro | custom
export RATE_LIMIT_STATE="/path/to/state.json"  # State file location
export RATE_LIMIT_WINDOW_HOURS="5"           # Rolling window duration
export RATE_LIMIT_ESTIMATE="200"             # Estimated request limit per window

Provider Presets

Presets are starting points. Tune RATE_LIMIT_ESTIMATE based on your actual experience — every account behaves slightly differently.

Tier System

Why 90 / 95 / 98?

These aren't arbitrary. Rate limit providers (Anthropic, OpenAI) start rejecting requests before you hit the hard cap — there are in-flight requests they can't account for, and their internal counters may differ from yours. The 90% threshold gives you a buffer to finish current work gracefully. By 95% you're in the danger zone where any burst could trigger a 429. At 98% you're one request away from a wall. The tiers create a smooth deceleration instead of a cliff.

Commands

python3 scripts/rate-limiter.py <command> [args]

gate                    # Check tier, exit code reflects severity
record [tokens]         # Log a request (tokens optional, default 0)
status                  # Full status JSON (tier, pct, requests, limit, backoff info)
pause [minutes]         # Enter paused state (auto backoff if no minutes given)
resume                  # Clear pause, reset to cautious
set-limit <n>           # Override estimated request limit
reset                   # Reset all state to defaults

Exit Codes

Complete Integration Example

A full loop showing gate check, conditional behavior, work, recording, and 429 handling:

#!/bin/bash
GATE=$(python3 scripts/rate-limiter.py gate 2>/dev/null)
EXIT=$?

if [ $EXIT -eq 2 ]; then
  echo "🛑 Critical/paused. Skipping work."
  exit 0
fi

if [ $EXIT -eq 1 ]; then
  echo "⚡ Throttled. Doing minimal work only."
  # skip sub-agents, background tasks, etc.
fi

# --- Do your actual work here ---
RESULT=$(your-agent-command 2>&1)

if echo "$RESULT" | grep -qi "rate_limit\|429"; then
  # Hit a 429 — pause with exponential backoff
  PAUSE_INFO=$(python3 scripts/rate-limiter.py pause)
  UNTIL=$(echo "$PAUSE_INFO" | python3 -c "import sys,json; print(json.load(sys.stdin).get('pausedUntil','unknown'))")
  echo "🛑 Rate limited. Paused until $UNTIL"
  exit 1
fi

# Record usage (estimate tokens based on your workload)
python3 scripts/rate-limiter.py record 2000

Agent Integration

In AGENTS.md / system prompt:

## Rate Limiting

Before expensive operations: `python3 scripts/rate-limiter.py gate`
- Exit 0 → proceed normally
- Exit 1 → reduce activity (no spawns, terse responses)
- Exit 2 → stop all non-essential work

After significant work: `python3 scripts/rate-limiter.py record <approx_tokens>`

On 429 error:
1. `python3 scripts/rate-limiter.py pause`
2. Stop current work
3. Set a timer/cron to run `python3 scripts/rate-limiter.py resume` at the pausedUntil time

In heartbeat checks:

## Rate Limit Gate (ALWAYS FIRST)
Run: `python3 scripts/rate-limiter.py gate`
- Exit 2 → reply HEARTBEAT_OK immediately. Do nothing else.
- Exit 1 → skip proactive checks. Only handle urgent items.
- Exit 0 → proceed normally.

In cron jobs:

Add to the start of any cron payload:

**FIRST: Rate limit gate check.** Run `python3 scripts/rate-limiter.py gate`.
If exit code is 2, reply 'RATE_LIMITED' and stop.
If exit code is 1, do only essential work.

How It Works

Agent → gate check → tier (ok/cautious/throttled/critical/paused) → adjust behavior
Agent → after work → record usage → updates rolling estimate
Agent → on 429 → auto-pause with exponential backoff → auto-resume

This skill uses heuristic estimation, not API-level usage data. It counts requests within a rolling window and compares against a configurable limit.

Why heuristic? Neither Anthropic nor OpenAI expose a real-time usage API. The usage pages (claude.ai/settings/usage, chatgpt.com/settings) require browser auth and scraping. This skill works out of the box with zero external dependencies.

Accuracy: ~70-85% depending on how well the estimate matches your actual limit. Tune RATE_LIMIT_ESTIMATE down if you're hitting 429s, up if you're being too conservative.

Improving accuracy:

• Start conservative (default presets)
• If you hit 429 → the skill auto-adjusts via exponential backoff
• After a few days, check status to see your actual request patterns
• Tune the estimate based on real data

State File

The skill writes a single JSON file (default: ./rate-limit-state.json). Structure:

{
  "provider": "claude",
  "plan": "max-5x",
  "tier": "ok",
  "estimatedPct": 23,
  "window": {
    "durationMs": 18000000,
    "requests": [{"ts": 1234567890, "tokens": 3000}],
    "estimatedLimit": 200
  },
  "backoff": {
    "consecutive429s": 0,
    "lastBackoffMs": 0
  },
  "pausedUntil": null
}

Why Not Just Handle 429s Manually?

The key difference: this is preventive, not reactive. Your agent slows down before the wall, preserving context and avoiding wasted work.

Troubleshooting

Hitting 429s despite ok status Your estimate is too high. Lower it: python3 scripts/rate-limiter.py set-limit 150 (or whatever feels right). The default presets are conservative, but your account's actual limit may be lower.

State file corrupted Reset everything: python3 scripts/rate-limiter.py reset. This clears all history and starts fresh. You won't lose configuration — just re-export your env vars.

Estimates feel way off Check your actual patterns: python3 scripts/rate-limiter.py status. Look at the request count vs. your limit. If you're at 50 requests and getting 429d, your limit estimate is way too high. If you're at 180/200 and never hitting limits, you can raise it.

Multiple OpenClaw instances Each instance needs its own state file. Set RATE_LIMIT_STATE to a unique path per instance:

export RATE_LIMIT_STATE="/path/to/instance-1-rate-limit.json"

Otherwise they'll overwrite each other's tracking and the estimates will be meaningless.

FAQ

What is this skill? Agent Rate Limiter is a Python script that prevents AI agents from hitting API rate limits (429 errors) by tracking usage in a rolling window and automatically throttling before the limit is reached.

What problem does it solve? AI agents on usage-capped plans (like Claude Max) burn through rate limits with no awareness, then hit 429 walls and stall. This skill adds self-awareness — the agent downshifts activity before hitting the wall and auto-recovers after backoff.

What are the requirements? Python 3 (standard library only). No pip installs, no API keys, no external services. Just a script and a JSON state file.

How does it work? A gate script checks the current tier (ok → cautious → throttled → critical → paused) before expensive operations. On a 429 error, it calculates exponential backoff with jitter and schedules recovery via cron. The agent reads the tier and adjusts behavior accordingly.

Does it work with any LLM provider? Yes. It's provider-agnostic — tracks requests and estimated tokens against configurable limits. Works with Claude, GPT, Gemini, or any API with rate limits.