agent-loop-guard
Framework-agnostic agent loop detection — sliding window similarity scoring to catch stuck agents.
Install
pip install agent-loop-guard
Quick Start
from loop_guard import LoopGuard, Action
guard = LoopGuard()
for action in agent_actions:
decision = guard.check(tool=action.name, args=action.args)
if decision.action == Action.STOP:
print(f"Loop detected: {decision.reason}")
break
Why not just max_iter?
| Approach | What it catches | Limitation |
|---|---|---|
max_iter=10 |
Runaway agents | Kills long legitimate tasks; misses 3-step loops at step 9 |
| agent-loop-guard | Exact repeats, fuzzy repeats, A→B→C→A cycles, output stagnation | — |
max_iter is a blunt timeout. agent-loop-guard detects behavioral patterns — the agent doing the same thing over and over, even with slight variations.
Detection Strategies
| Strategy | What it detects | Confidence signal |
|---|---|---|
| Exact Repeat | Same (tool, args) called repeatedly |
Consecutive identical calls |
| Fuzzy Repeat | Near-identical args (Jaccard + edit distance) | Similarity > threshold |
| Cycle Detection | A→B→C→A→B→C repeating sequences | Pattern repetition count |
| Output Stagnation | Tool returns same output repeatedly | Output similarity > threshold |
All four strategies run on every call. The highest confidence wins.
API
guard = LoopGuard(
window_size=10, # actions to keep in memory
similarity_threshold=0.85, # fuzzy match threshold
)
decision = guard.check(
tool="web_search", # tool/function name
args={"query": "python"}, # arguments (dict or str)
output="Results: ...", # optional: enables stagnation detection
)
decision.action # Action.CONTINUE | WARN | STOP | ESCALATE
decision.reason # "Cycle detected: [search → parse → search] repeated 3 times"
decision.strategy # "cycle_detection"
decision.confidence # 0.0 ~ 1.0
decision.is_loop # True if STOP or ESCALATE
decision.should_warn # True if WARN
guard.reset() # reuse for next session
Action Escalation
Actions escalate with consecutive detections:
from loop_guard import ActionConfig
config = ActionConfig(
warn_threshold=2, # 2 consecutive hits → WARN
stop_threshold=4, # 4 consecutive hits → STOP
escalate_threshold=6, # 6 consecutive hits → ESCALATE
)
guard = LoopGuard(action_config=config)
Generic Callback
from loop_guard.integrations.generic import LoopGuardCallback
callback = LoopGuardCallback(
on_warn=lambda d: logger.warning(f"Loop warning: {d.reason}"),
on_stop=lambda d: raise_stop_error(d),
)
# In your agent loop:
decision = callback.before_tool_call("search", {"query": "test"})
License
MIT
Release files for agent-loop-guard 0.1.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| agent_loop_guard-0.1.1.tar.gz | 16.2 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| agent_loop_guard-0.1.1-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 26.5 kB
Release files / agent_loop_guard-0.1.1.tar.gz
| Download URL | agent_loop_guard-0.1.1.tar.gz |
|---|---|
| Size | 16.2 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
52cc8c0572509696b073d79108a010be9fcce32ada8354f92c9900866073b07c
|
|
BLAKE2b-256 checksum How to use checksums |
3349da649b8f53e0684e77a053b17b3f184f82281e9d5cd98611cbe3e3641ebe
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/6.1.0 CPython/3.13.7
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Mar 25, 2026.
Transparency logRelease files / agent_loop_guard-0.1.1-py3-none-any.whl
| Download URL | agent_loop_guard-0.1.1-py3-none-any.whl |
|---|---|
| Size | 10.4 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
5f5d487f39f04120a44c3f0ab3d02a4ccaa130c6c8361e79e7b21f83029bb52a
|
|
BLAKE2b-256 checksum How to use checksums |
f1e43ab6c5a071297633f101f64c94d09583332df564e2f22575b00400a739cc
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/6.1.0 CPython/3.13.7
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Mar 25, 2026.
Transparency log