Southbridge provides an open‑source runtime called Hankweave that lets teams build, run, and maintain long‑horizon AI agents—called “hanks”—for complex data ingestion and transformation tasks. The platform uses modular codon blocks, preflight validation, checkpointing, and antibrittle design patterns to keep agents reliable and cost‑aware over hundreds of hours of execution, enabling enterprises and research labs to unlock siloed data at scale.
Funding
Funding not disclosed
Founders
Product
Problem
Organizations accumulate massive amounts of data in heterogeneous, legacy schemas and siloed pipelines, making it difficult to ingest, unify, and continuously process information at scale. Traditional LLM‑driven agents break down over long executions, leading to brittle, unreliable automation for complex data workflows.
Solution
Southbridge offers a runtime called Hankweave that enables the creation, execution, and maintenance of long‑horizon, repairable AI agents—referred to as “hanks”—for data‑intensive tasks. By structuring agent logic into reusable codons with built‑in preflight validation, checkpointing, and context‑boundary controls, Hankweave ensures agents can run for hundreds of hours and thousands of tool calls without failure. The platform adds “antibrittle” design patterns such as sentinels, controlled forgetting, and structured event journals to detect and correct drift, laziness, and other failure modes automatically. All components are open‑source, allowing teams to integrate the runtime with existing LLM harnesses (e.g., Claude Agent SDK, Codex) and to share repaired codons across projects, reducing duplicated engineering effort. This approach unlocks first‑mile data ingestion and transformation at enterprise scale while keeping the workflow transparent, auditable, and cost‑aware.
Target Audience
Primary customers are data‑heavy enterprises, AI research labs, and engineering teams that need reliable, long‑running autonomous agents for data ingestion, transformation, and analysis across domains such as finance, healthcare, genomics, and insurance.
Features
- Codon‑based modular agent blocks that encapsulate prompts, workspaces, and behavior checks, reusable across projects
- Preflight validation of API keys, model availability, file paths, and configuration to prevent runtime failures
- Checkpointing, rollback, and archive mechanisms that enable safe long‑running executions and easy recovery
- Sentinels and structured event journals for real‑time detection of brittleness, drift, and policy violations
- Controlled forgetting and explicit context handoffs to manage memory and maintain coherence over extended runs
- Compatibility with major LLM agent harnesses (Claude, Codex) and a unified, cost‑aware execution interface
- Open‑source runtime (Hankweave) with documentation, examples, and community contributions