Sign InOpen Brain
AI EngineerVideoSource Linked

Building Agents Is Trivial Now, Context Is the Next Frontier — Jeff Ng, Unblocked

A coding agent with repository access still recommended a known outage trigger. Background agents need reconciled decisions and incident context, not just more raw MCP results.

AI Engineer · Aug 21, 2026
Open Source Open MarkdownOpen JSON
Source Summary

An agent investigating QA latency of **3–4 seconds** versus an expected **hundreds of milliseconds** recommended re-enabling async dispatch. It missed a later Slack discussion and Linear postmortem showing that setting had caused an outage.

Practical Implication

For unattended coding agents, retrieve decisions, incidents, tickets, conversations, and code together. Reconcile conflicts, rank relevance, enforce permissions, and provide a synthesized task view rather than making each agent resolve raw MCP output.

Agent-Ready Context
An agent investigating QA latency of **3–4 seconds** versus an expected **hundreds of milliseconds** recommended re-enabling async dispatch. It missed a later Slack discussion and Linear postmortem showing that setting had caused an outage.

For unattended coding agents, retrieve decisions, incidents, tickets, conversations, and code together. Reconcile conflicts, rank relevance, enforce permissions, and provide a synthesized task view rather than making each agent resolve raw MCP output.

The example demonstrates one corrected recommendation, not a measured reliability gain. Building and maintaining an organization-wide context model also introduces access-control, freshness, and conflict-resolution work.
Connected Context · Feed7 Judgment

This broadens coding-agent retrieval from repository evidence to organizational decision history: technically plausible advice can be wrong when later incidents, tickets, or conversations reverse the relevant context. It therefore raises synthesis, conflict reconciliation, freshness, and permissions to first-class context-layer responsibilities, though the single corrected case does not quantify reliability gains.

infiniflow/ragflowRAGFlow offers a self-hostable substrate for mixed private sources and citations, but this signal adds the need to reconcile conflicting decisions and produce a permission-aware task view.AI on Your Lakehouse: Context Comes in Shapes, Not Queries — Zach Blumenfeld, Neo4jGraph-shaped retrieval could represent links among code, incidents, tickets, and decisions that isolated queries miss, subject to local evaluation.tirth8205/code-review-graphcode-review-graph narrows coding context through dependency evidence; this signal shows that code relationships alone can still omit the organizational history that determines whether a change is safe.RLM: Recursive Language Models for Large Codebases - Shashi, Superagentic AIRLM-style bounded inspection helps control large repository context, but the outage example shows that the external evidence surface must extend beyond the repository.
Context Map
contextcoding#context-engineering#retrieval#mcp
Uncertainty
The example demonstrates one corrected recommendation, not a measured reliability gain. Building and maintaining an organization-wide context model also introduces access-control, freshness, and conflict-resolution work.