Reliability and evaluation
Treat control-context compression as a behaviour change
Research preprint · 2 August 2026
The authors of Control Under Compression evaluate compressed system prompts through executable agent outcomes. Their results vary by context and compression method. For your system, the useful question is which tool behaviour changes when those prompts are shortened.
Examine what happens at the compaction boundary
Research preprint · first submitted 6 August 2026; revised 14 October 2026
Min and colleagues study execution instability after repeated context compression. Their TRACE approach evaluates compaction through paired continuations. This is early empirical evidence; it suggests testing the behaviour following compaction rather than judging only the quality of a summary.
Security and evidence
Include poisoned data in the threat model
Research preprint · July 2026
The authors of Agent Data Injection Attacks are Realistic Threats to AI Agents examine attacks on the data an agent uses. The review question is how your system checks retrieved facts before an action depends on them.
Check the authority behind consequential actions
Vendor security research · Zscaler ThreatLabz
ThreatLabz describes indirect prompt injection in web content aimed at AI agents. Its sandboxed test conditions matter when interpreting the results. For a design review, examine tool permissions, spending boundaries and approval for consequential actions alongside the model prompt.
Protocol and deployment
Read the MCP specification and migration guidance together
Official project publications · 2026-07-28 specification
The Model Context Protocol project describes changes to its protocol and SDKs. Review the version your client and server actually negotiate, the migration requirements and the effect on your deployment architecture. Check the current project guidance before applying a change.
