Skip to content

Interlens: Framework for Multi-Agent Interaction and Interpretability

This library provides a harness, optimized utilities, and interpretability hooks for multi-agent conversation rollouts.

A harness for multi-agent (model-to-model) conversations with first-class interpretability—activation capture, steering, activation patching, and token logprobs—all hooked into the same generation path as real turns and tagged to conversation structure. Scales from one interactive dialogue to thousands of checkpointed, multi-GPU rollouts.

from interlens import Conversation

conv = Conversation.from_models(
    ("Qwen/Qwen2.5-0.5B-Instruct", "Qwen/Qwen2.5-0.5B-Instruct"), names=("alice", "bob"),
    shared_context="Let's debate: is cereal a soup?",
)
conv.run(turns=4, first="alice")
print(conv.transcript)

See docs/examples for sample code.

Install

pip install interlens
# with hosted-API participants (APIParticipant):
pip install "interlens[api]"

PyTorch / CUDA note

torch is declared as a plain, build-agnostic dependency — install the wheel matching your platform (CUDA / CPU / MPS) before or alongside interlens. E.g. for CUDA 13.0: 

pip install torch --index-url https://download.pytorch.org/whl/cu130
See https://pytorch.org/get-started/locally/.

What's inside

  • Conversation — turn-taking over a shared, perspective-neutral Transcript; per-speaker view pipeline (system/private framing → context-fit → family-correct chat template).
  • AutoModelParticipant — HF-style factory (from_pretrained / from_model / from_) that returns the family-correct participant (Qwen/Gemma/…); APIParticipant for hosted models.
  • Interpretability — conv.capture(...), SteeringSpec, Patch, token_logprobs, backed by a queryable ActivationCache.
  • Scale — conv.rollout(...) / interlens.run([...]): multi-GPU, checkpointed, resumable, batched co-stepping, with in-worker analyzer callbacks; data-driven rollouts via dataset_field, matched compute via TokenBudget.
  • One object, no ceremony — a Conversation (with lazy participants) is at once the serializable recipe, the live dialogue, and the rollout driver; build it functionally (.turns(6).data(ds).analyzer(grade)), .set(...) copy-on-write, and save/load (recipe + transcript).

See docs/examples/ for a simple→advanced walkthrough of the whole API.

Develop

git clone https://github.com/Sid-MB/interlens && cd interlens
uv sync                     # installs the package + dev group (pytest, pre-commit)
uv run pre-commit install   # one-time: activate the AGPLv3 license-header git hook
uv run pytest               
# fast tests; opt-in to thorough tests requiring downloading models + a GPU with: pytest -m slow

License

GNU AGPLv3 — see LICENSE.