Every processed story in chronological order, with the newest coverage first. Filter by tag, source, or score to drill in.
The release extends the E2B sandbox filesystem API with persistent, queryable file metadata and richer filesystem events, while closing a broad set of behavioral inconsistencies between the JS and Python SDKs.
The system directly addresses the structural reason Claude Code sessions lose productivity — no persistent project memory — by encoding context in `CLAUDE.md` and enforcing workflow discipline that keeps every session starting with full context and every change safely reversible.
The research reframes where agent cost optimization efforts should focus — not on code generation, but on the iterative code review loop, where a structural "communication tax" drives the majority of token spend.
The adaptive tool loading profile cuts per-session token consumption by allowing clients to load as few as 13 of the plugin's 43 tools, while the excluded-files filter closes a gap where sensitive or irrelevant vault content could surface in semantic search results.
The mid-execution `ask_user()` mechanism allows agentic tools to gate side effects on explicit human approval and survive server restarts while awaiting a response, replacing a model where agents had to complete or abort a turn without user input.
Using Claude's tool-calling with a strict `input_schema` eliminates the markdown-fence JSON parsing failure mode that plagues free-text LLM output, making AI-generated config files reliably writable to disk without a fragile `JSON.parse` step.
The paper fills a documented gap by writing down, for the first time in a consolidated form, the end-to-end practice for building production custom AI agents — knowledge the authors note has previously existed only in informal sources like podcasts, blogs, and leaked system prompts.
Hades replaces token-heavy YAML parsing with a structured MCP graph query layer, directly addressing the missed-dependency errors that stock Claude Code produces when reasoning about Unity project relationships.
The paper demonstrates that replacing linear repository traversal with domain-scoped parallel agent spawning improves multi-file change localization for a small model, while also identifying that naive filesystem access and forced multi-agent consultation can actively harm performance or inflate costs.
The benchmark reveals that frontier coding agents can reliably execute computational social science workflows, while also exposing prompt-framing vulnerabilities that could introduce bias into AI-assisted scientific production.