Sleuth: A Structured AI Detective Workflow Framework, Building a Complete Investigation Pipeline with 56 Tasks

Sleuth is an open-source six-stage AI agent investigation workflow framework. It achieves structured reasoning similar to intelligence analysis through 56 fine-grained tasks and templated processes, solving problems such as broken reasoning chains and information integration difficulties for AI agents in complex tasks. It provides a complete investigation pipeline from task orientation to final report output, suitable for various in-depth investigation scenarios.

Currently, AI agents often face problems like broken reasoning chains and information integration difficulties in complex tasks, especially in multi-step investigation and cross-document correlation analysis scenarios. The lack of a systematic workflow leads to unstable output quality. The structured methodologies accumulated over a long time in the intelligence analysis field can provide reusable frameworks for AI agents' complex reasoning.

Sleuth is an open-source project developed by elb-pr, providing a six-stage AI agent investigation workflow framework. It decomposes complex investigation tasks into 56 fine-grained tasks, with clear input/output specifications and reference templates for each stage. The core execution engine is task_runner.py, which is responsible for driving the completion of tasks in each stage in order.

1. Task Orientation: Clearly define the investigation objectives, scope boundaries, and success criteria; structurally define key questions, stakeholders, and information source priorities.
2. Intelligence Collection: Large-scale information collection + source credibility assessment + collection strategy optimization; provides multiple collection templates and a source quality scoring mechanism.
3. Collation and Entity Resolution: Identify key entities (people/organizations/events/locations) and resolve disambiguation issues; output a cleaned and linked entity graph.
4. Timeline and Relationship Processing: Build timelines and relationship networks; output visual timeline diagrams and relationship graphs.
5. Hypothesis and Reasoning: Generate competitive hypotheses and collect supporting/refuting evidence to reduce confirmation bias.
6. Final Report: Summarize results, including executive summary, methodology, key findings, evidence chain, confidence assessment, and recommendations; standardized format for easy review and comparison.

task_runner.py implements automatic flow between stages and state management. Each completed task outputs reference files for downstream use, forming a clear data dependency chain. This design makes long processes traceable and recoverable—if a stage fails, it can resume from the breakpoint instead of starting over.

The Sleuth framework is suitable for academic research literature reviews, corporate competitive intelligence analysis, open-source intelligence investigations, codebase architecture sorting, complex customer support ticket processing, etc. Any task that needs to extract structured insights from scattered information can benefit from it.

Sleuth transforms mature methodologies in the intelligence analysis field into executable workflows for AI agents, filling the structural gap in current AI tools for complex reasoning tasks. The 56 fine-grained tasks ensure both process integrity and flexibility, making it a reference implementation worth in-depth study for developers building reliable AI investigation systems.