Interlock: Moving Long-Horizon Agents from Local Correctness to Task Completion

Introduction

A long-horizon agent can fail even when many of its individual actions look reasonable. It may find the right evidence, make a sensible decision, or edit the expected file, yet still leave the overall task incomplete. The missing piece is often not a single capability, but the relationship between evidence, decisions, and their downstream effects.

Interlock targets this gap. The paper argues that task synthesis should not stop at producing a natural-language prompt. A useful task must also make its dependencies explicit, expose the evidence available to the solver, and define how completion will be verified.

A graph for completion

The framework represents each task as a typed completion graph. The graph captures more than a sequence of actions; it describes the conditions that must hold together:

• Grounding edges connect public environmental evidence to the decisions an agent is expected to make.
• Coupling hyperedges impose joint constraints across several components, covering requirements that cannot be evaluated independently.
• Realization edges connect decisions to every required downstream effect, such as changes to artifacts or environment state.

Interlock compiles this structure into three coordinated outputs: a resettable workspace, a public request, and an evaluation rubric. This makes the environment, the instruction, and the acceptance criteria parts of one task specification rather than separate artifacts assembled independently.

Checking whether a task is actually testable

The generation pipeline includes a source-coverage review to determine whether the original requirements survive task construction. It also performs a public-input review, asking whether the rubric is supported by information the solver can actually access. This distinction matters. An evaluator that relies on hidden evidence can reject a reasonable trajectory for reasons the agent could never have anticipated.

After execution, an agent-as-judge inspects trajectories and resulting artifacts, while verification scripts provide additional checks. Only trajectories judged to satisfy all mandatory requirements are retained for training. The authors fine-tune two Qwen3.6 backbones on these accepted trajectories. According to the supplied abstract, AutomationBench pass rates improve by 4.00–4.50 percentage points. The provided material does not include the full results for the other benchmark mentioned in the abstract, so broader performance claims should be treated cautiously.

Why it matters

Interlock reframes task completion as a relational property. The question is not simply whether an agent took plausible actions, but whether evidence supported the decisions and whether those decisions produced every required effect. This is particularly relevant to office automation, tool-use workflows, and stateful environments where a missing update or broken dependency can invalidate an otherwise strong trajectory.

The approach also highlights a path beyond scaling models alone. Better task specifications, evidence boundaries, and artifact-level verification can improve the quality of data used to train agents. At the same time, the framework introduces practical challenges: constructing completion graphs may require substantial effort, judge decisions can be imperfect, and verification scripts may not cover every failure mode.

Interlock is therefore best understood as a task-engineering and acceptance protocol for long-horizon agents. As agents take on more cross-application work, making dependencies explicit may become as important as improving the model’s reasoning ability.

Source: Hugging Face Daily Papers