Agent-runtime review targets

Candidate runtimes to review once topology and grain are settled (see README.md). A target here is registered, not yet scheduled — the review type/skill and collection are still open decisions.

Queue

1. flue — withastro/flue

  • Repo: https://github.com/withastro/flue
  • Self-description: "The sandbox agent framework."
  • Language: TypeScript
  • Why a runtime, not a memory system: describes itself as an agent framework built around a sandbox — i.e. the execution-substrate + scheduler layers, with memory (if any) as a subsystem. Good first probe of whether the Scheduler / Context-engine / Execution-substrate skeleton holds against a real runtime.
  • Status: registered (first target). Not yet reviewed.
  • Captured: 2026-06-06

2. cocoindex — cocoindex-io/cocoindex

  • Repo: https://github.com/cocoindex-io/cocoindex
  • Self-description: "Incremental engine for long horizon agents."
  • Language: Rust (Python bindings)
  • Topics: context-engineering, indexing, rag, data-indexing, change-data-capture, knowledge-graph, semantic-search, long-horizon-agent, agentic-data-framework
  • Boundary flag: likely a Context-engine / Execution-substrate subsystem (incremental indexing, CDC, RAG pipeline) rather than a full scheduler-bearing runtime — it feeds context in, it doesn't run the loop. Good test of the grain/boundary decision: does a runtime subsystem get its own review, slot into a runtime review, or sit closer to the existing memory corpus? Contrast with flue, which is execution-substrate + scheduler.
  • Status: registered. Not yet reviewed.
  • Captured: 2026-06-06