Four Bets to Fix Agent Stack Ceilings
Production agents fail due to governance gaps from shared credentials, siloed context, fragile sessions, and custom plumbing—bet on platform-level identities, universal context, durable execution, and open platforms.
Embed Identities at Platform Level for Governance
Agents today borrow shared service accounts or OAuth tokens, creating governance debt: zero visibility into tool access, data movement, or credentials used. This leads to runtime failures like schema drifts or PII leaks, with agents marking tasks complete amid corrupted state. Fix by moving agent identity to the platform layer—distinct, unforgeable identities enforced at network level, like a badge reader. If an agent accesses an uncleared database, the connection blocks instantly. Result: attributable actions, auditable permissions, revocable agents, turning liabilities into a managed workforce without latency-heavy middleware.
Deliver Universal Context and Mission Persistence
Context is siloed—browser agents see tabs, desktop ones see files—but business systems (CRM, ERP, data warehouse) remain inaccessible across tasks, wasting engineering on custom serialization and memory layers. Agents forget missions mid-task. Solution: platform-integrated universal context spanning all systems simultaneously, enabling reasoning over real business data.
Sessions die on laptop closes or token limits, unfit for enterprise workflows like week-long procurement or month-long audits. Achieve durable execution with cloud-native state/checkpointing surviving restarts, disconnects, redeploys, and model changes (no local Redis hacks). Use long-horizon memory, summarization, and handoffs for multi-week tasks. Support missions outliving sessions via credential rotations and audit trails. Add first-class human-in-the-loop pauses for approvals. This raises the bar from hour-long demos to quarter-long autonomy with guardrails.
Prototype on Open Platforms, Not Custom Plumbing
Teams waste cycles on undifferentiated work: custom memory, eval harnesses, observability, retries—draining bandwidth from domain-specific logic like company judgment calls or regulations. Copy cloud/container model: prototype locally on open-source orchestration primitives, scale to managed platforms without rewrites. This lets engineers focus on product-differentiating business logic, not boilerplate.
Over five years, winners choose agent foundations covering identity, context, persistence, orchestration—freeing months for unique problems. Rebuilding stack monthly delays deployment; build on agent-engineered platforms instead.