Updated
Updated · MIT Technology Review · Jul 27
Outshift Unveils 3 Open-Source AI Layers, Lifting Agent Coordination to 93%
Updated
Updated · MIT Technology Review · Jul 27

Outshift Unveils 3 Open-Source AI Layers, Lifting Agent Coordination to 93%

2 articles · Updated · MIT Technology Review · Jul 27

Summary

  • Outshift by Cisco said its new open-source stack—AGNTCY, Mycelium and CASA—aims to let AI agents share intent, context and reasoning so they can coordinate decisions without human stitching.
  • Mycelium is the semantic coordination layer: in internal tests across 14 scenarios, structured agent groups reached decisions 93% of the time versus about one-third for unstructured groups.
  • AGNTCY provides the underlying connectivity layer for agent discovery, identity and messaging across systems and companies, extending a Linux Foundation project that already includes more than 75 participants.
  • CASA adds continuous task-level authorization to curb risks such as prompt injection, memory poisoning and over-privileged agents; Outshift said agents today cannot confirm they are cleared for a job roughly 90% of the time.
  • The push reflects a broader shift from scaling single models vertically to scaling intelligence horizontally across enterprises, where current open-source multi-agent systems still show failure rates of 41% to 87%.

Insights

As AI agents break out of internal silos, will the heavy security needed for cross-company trust destroy their real-time speed?
Can cryptographic identities truly stop malicious AI agents from secretly poisoning cross-company workflows before human overseers notice?
If an autonomous AI negotiates a disastrous deal with a partner's compromised bot, who legally pays the price?