able-harness-kit

Thin, backend-neutral agent-harness middlewares for LangChain / deepagents.

Production-distilled patterns that compose on top of any LangChain AgentMiddleware stack (including deepagents). They don't replace your agent — they harden the loop around it.

pip install able-harness-kit          # after the first PyPI release
# or, from a clone:
pip install -e .

Why

Most agent failures aren't "the model isn't smart enough" — they're the loop around the model lacking observation and control:

Each is backend-neutral (depends only on LangChain / LangGraph public types), small enough to read in one sitting, and composes with — rather than replaces — your existing harness.

Usage

from langchain.agents import create_agent          # or deepagents.create_deep_agent
from able_harness_kit import (
    ProgressAwareLoopGuardMiddleware,
    BinaryReadGuardMiddleware,
    ToolResultBudgetMiddleware,
)

agent = create_agent(
    model,
    tools=[...],
    middleware=[
        BinaryReadGuardMiddleware(),                   # fail loud on binary read_file
        ToolResultBudgetMiddleware(limit=16_000),      # cap oversized tool results
        ProgressAwareLoopGuardMiddleware(stop_at=3),   # stop loops that make no progress
    ],
)

ProgressAwareLoopGuardMiddleware

The cheap loop signal — "same tool, same args, N times" — is wrong about half the time. Polling a job, walking pagination, or tailing a stream all repeat the call and should: each one returns new information. The difference between progress and a stuck loop isn't in the arguments, it's in the output.

This guard runs the tool, compares the result to the previous same-tool result, and intervenes only when a call both repeats and returns no new information (same_args_repeat_count >= stop_at and new_information_delta < progress_floor). Genuine progress passes through; a true spin gets a model-facing directive (not a synthetic user message). For tools where a content-hash comparison is too lax — a TODO tool re-sent with a one-character cosmetic edit, say — pass delta_overrides={"write_todos": make_set_membership_delta("todos")}.

The output-delta primitives are exported standalone: build_observations(state["messages"]) reconstructs the per-call observations (ToolCallObservation with new_information_delta) from history with no instance state.

LoopGuardMiddleware

The cheaper variant: fingerprints each call (name + normalized args), counts consecutive identical-argument repeats, and short-circuits once stop_at is hit — without looking at the result. Use it when you only want to catch a model hammering an identical call and don't need output tracking. The counter resets the moment a different call is seen. Pass on_signal= to observe without changing control flow.

BinaryReadGuardMiddleware

Intercepts read_file results whose declared media type isn't text-like and replaces the base64 payload with a structured error pointing at a dedicated extractor (OCR / doc-to-markdown / type sniff). Turning a silent hallucination into an explicit "use the right tool" nudge is strictly safer; the only cost is one extra tool hop on the first attempt. Tool name, metadata keys, and the directive are configurable; disable per-deployment with BINARY_READ_GUARD_ENABLED=0.

ToolResultBudgetMiddleware

Caps each tool result at a character budget. Oversized results are offloaded via a caller-supplied offload(content, tool_name) -> OffloadRef callback (bring your own store — filesystem, blob, vector DB), or truncated with a retrieval hint by default.

Background

These were distilled from a multi-agent product running on non-Anthropic model gateways (Qwen / GLM / DeepSeek), and from a controlled experiment across 5 models × 9 behavioral probes (harness on vs. off) that framed a harness as an entropy-reducing control system rather than overhead. The progress-aware loop signal — keying off whether a repeated call returned new information, not just whether the arguments repeated — is the open distillation of that controller's loop detector. The middlewares here are the reusable, framework-neutral core of that work.

License

MIT © Hanlin Li