Digest #1 · 2026-10-13

Cross-checking agent code, sandboxing agents, and RAG without a vector DB

Spotlight: Testing & evals. This digest leans on verification: independent review of agent-written code, and an eval-minded view of whether agents actually ship better software.

Top pick to test this week

OpenQodex: pre-push code review by a different agent

Testing & evalsHelios30 min

Runs deterministic scanners (gitleaks, semgrep) on your unpushed changes, then starts a separate Claude Code or Codex process as reviewer. By default the agent that wrote the code does not review it. Runs locally on your own login; Apache-2.0.

Cheap second-opinion pattern for agent-written code, and the 'scanners first, model second' split is a design you can copy into Helios gates.

Test this: Install it, then run it on the next uncommitted Helios change. Compare its findings with your own review. 30 min.

Tools & repos

Strands Box

Agents1 hr

Open-source sandbox for agents: OS-level isolation plus semantic access policies, configured in box.toml and policy.dw. Works with any agent binary.

Gives you least-privilege for agents that touch client files or Gmail, instead of trusting prompts.

Test this: Wrap a throwaway Claude Code session around a copy of southport/ and see what you can and cannot restrict.

VectifyAI jev-doc-search + PageIndex

RAG & context45 min

Long-document search with no vector DB or embeddings: the document becomes a hierarchical tree and the model walks it choosing among a few options per level. Needs a Jev API key.

A concrete alternative to chunk+embed RAG for long structured docs (contracts, reports, handbooks).

Test this: Run page_search.py on a long PDF you own and compare with a plain embedding search. Note the API key requirement.

IHMT-MEMORY

RAG & context15 min

Long-term memory for coding agents (Claude Code, Codex, opencode). Early project, 30 stars at collection time.

Memory layers are an active area; worth a look against your current memory-file approach, but treat as unproven.

Test this: Read the README and decide if it beats CLAUDE.md plus memory files before installing.

Talorys

Automation20 min

Self-hosted personal AI agent that runs on Cloudflare's free tier (Show HN, 300+ points).

Cheap always-on agent hosting model; relevant to running your own scheduled agents without a Mac awake.

Test this: Skim the architecture and note what a free-tier always-on agent can and cannot do.

Ways of working

Build a retrieval layer for code, not just a bigger context window

RAG & context

JetBrains' field notes (part 1) on parsing, chunking and vectorizing repos so agents get citable evidence instead of whatever grep surfaces.

Same problem as pointing an agent at Helios or the vault. Chunk by code structure, not size.

Test this: List where your agents currently rely on grep over notes; pick one corpus (southport/notes) to prototype retrieval on.

Harness engineering is the new craft

Agents

Writeup on engineering an AI harness for game development with Claude Code.

Shows the loop of tools, checks and feedback around the agent, which matters more than the prompt.

Test this: Identify one repetitive Helios task and write down the verification step a harness would need.

Claude Code's suggested-message feature

Refinement

Opinion piece (the original blog post now 404s, so this links the 269-point HN discussion) arguing the suggested-message feature really serves the model.

A useful lens on how products steer agent behavior; skim for ideas, it is commentary rather than a how-to.

Test this: Read for 5 minutes; no setup needed.

Company use cases

JetBrains

Built Air Context, a RAG pipeline giving LLM agents precise, citable code evidence, and published the lessons from production.

Steal this: Treat retrieval quality as a measured component with its own tests, separate from the agent.

Wikimedia Foundation

Reports unauthorized agent activity found on Wikimedia projects (307 HN points). A cautionary use case rather than a success story.

Steal this: Agents that act on the outside world need identity, rate limits and audit trails. See Strands Box above.

Reading list

Sample digest #1, built from live HN and GitHub data on 2026-10-13. Web search returned little October-dated material, so this digest leans on the collector. Items were summarised from READMEs and page text; HN scores are at collection time.