Releases·Tool Release
Your Coding Agent Is 6 Months Out of Date — Jakub Hojsan, Exa
In a talk recorded at AI Engineer World's Fair 2026, Exa forward deployed engineer Jakub Hojsan explains how models ship roughly six months behind their knowledge cutoff, leaving code-review agents unable to assess recent changes. He presents Exa's approach: agent rules for when to search, plus query-dependent highlights that compress a 100,000-character page to about 500 characters without added latency. The talk matters to practitioners building agents on top of models with stale training data, and previews Exa Agent.
23 engagement1 sourceFri, Oct 2, 2026, 03:30 PM
AI Overview
- Exa FDE Jakub Hojsan says models ship ~6 months behind their knowledge cutoff, leaving code-review agents reviewing changes they've never seen (AI Engineer World's Fair 2026).
- Hojsan argues adding a web search tool alone doesn't fix the gap; agents need explicit rules for when to search.
- Exa's query-dependent highlights cut a 100,000-character page to ~500 characters with no added latency, previewing Exa Agent and provider-independent search via Exa MCP.
Related
Social·Community
Peter Werry (Unblocked) argues code agents mistake search results for codebase understanding
In a YouTube talk, Unblocked's Peter Werry describes how coding agents fail on real codebases: they search, find one plausible answer, and stop — a failure mode he likens to radiologists' 'satisfaction of search.' He argues access to information is not understanding and frames context engines as the fix, replacing the human role of carrying tribal knowledge.
In a YouTube talk, Unblocked's Peter Werry describes how coding agents fail on real codebases: they search, find one plausible answer, and stop — a failure mode he likens to radiologists' 'satisfaction of search.' He argues access to information is not understanding and frames context engines as the fix, replacing the human role of carrying tribal knowledge.
22 engagement1 sourceyoutubeThu, Aug 27, 2026, 06:00 PM
Developments·Paper
OpenCodeReview: Determinism over Non-Determinism for Cost-Effective Agent-Based Code Review
A new arXiv paper introduces OpenCodeReview, an agent-based code review system that prioritizes deterministic engineering to address non-determinism and context locality in LLM-based reviewers. The approach targets unstable outcomes, bounded diff context, and hallucinated comments, aiming for cost-effective, scalable review.
A new arXiv paper introduces OpenCodeReview, an agent-based code review system that prioritizes deterministic engineering to address non-determinism and context locality in LLM-based reviewers. The approach targets unstable outcomes, bounded diff context, and hallucinated comments, aiming for cost-effective, scalable review.
0 engagement1 sourcearxivMon, Aug 10, 2026, 08:43 AM
Releases·Tool Release
Silica-core: "No-Model-in-the-Loop" MCP retrieval engine for coding agents cuts tool calls 28%
A developer building Silica-core, an open-source MIT-licensed retrieval engine exposed over MCP and CLI, describes its design constraint: the harness owns the loop. The tool returns only located evidence (file path, section, line window, mathematical signals) and never summarizes or decides for the model, a pattern the author reports cut tool calls by 28% with coding agents like Claude Code, Cursor, OpenHands, and Zed.
A developer building Silica-core, an open-source MIT-licensed retrieval engine exposed over MCP and CLI, describes its design constraint: the harness owns the loop. The tool returns only located evidence (file path, section, line window, mathematical signals) and never summarizes or decides for the model, a pattern the author reports cut tool calls by 28% with coding agents like Claude Code, Cursor, OpenHands, and Zed.
1 engagement1 sourceredditWed, Sep 9, 2026, 03:11 PM
Developments·Paper
CodeGrep: RL-trained retrieval agent cuts coding-agent token waste
A new paper introduces CodeGrep, a 14B retrieval agent trained with GRPO to handle repository exploration for LLM coding agents. It aims to reduce the token overhead spent on grep, glob, and file reads, which currently dominate coding-agent budgets.
A new paper introduces CodeGrep, a 14B retrieval agent trained with GRPO to handle repository exploration for LLM coding agents. It aims to reduce the token overhead spent on grep, glob, and file reads, which currently dominate coding-agent budgets.
0 engagement1 sourcearxivThu, Aug 6, 2026, 11:07 AM
Products·building
AI Codebase Memory Agent: persistent knowledge layer for repositories
An AI agent that ingests a codebase and builds a persistent memory of development knowledge, then answers questions with evidence. LLMs extract useful knowledge (decisions, docs, PRs) during ingestion and power retrieval. It targets developers working in unfamiliar codebases who otherwise repeatedly search files and history to understand why things exist.
An AI agent that ingests a codebase and builds a persistent memory of development knowledge, then answers questions with evidence. LLMs extract useful knowledge (decisions, docs, PRs) during ingestion and power retrieval. It targets developers working in unfamiliar codebases who otherwise repeatedly search files and history to understand why things exist.
2 engagement1 sourceredditTue, Sep 29, 2026, 04:22 PM