Analyst Urges 4 Architectural Bets to Fix Agentic Systems’ Governance and Reliability Gaps
Updated
Updated · O'Reilly Media · May 20
Analyst Urges 4 Architectural Bets to Fix Agentic Systems’ Governance and Reliability Gaps
7 articles · Updated · O'Reilly Media · May 20
Four architectural changes should become priorities within 12 months, the analyst argues, because today’s production agents still rely on fragile plumbing, broad permissions and weak audit controls.
The first fix is platform-level identity: agents should have distinct, enforceable credentials instead of shared service accounts or borrowed human tokens, making every action attributable and revocable.
The second and third bets target execution limits—universal context across business systems and durable, cloud-native persistence so multi-step work can survive disconnects, handoffs, credential rotations and long-running tasks.
The fourth bet is to build on open agent platforms rather than custom memory, observability and retry stacks, shifting engineering time from undifferentiated infrastructure to domain-specific business logic.
Over a 5-year horizon, the piece argues, teams that standardize identity, context, persistence and orchestration will outpace those still rebuilding the agent stack from scratch.
Are complex new AI governance platforms a critical safety measure or just innovation-killing bureaucracy in disguise?
When an AI agent with a unique identity causes a disaster, who is ultimately held accountable?
Will the Shopify-Google AI commerce protocol create a new tech monopoly or level the playing field for small businesses?