Solenne Kaboré

Scheduled agent fleets need four separate recovery contracts, not just retries.

Smaller teams need plug-and-play platforms.

Companies need to track workflow-level costs and outcomes to prove AI's profit impact.

Agents need one runtime that owns state, memory, and scheduling.

How semantic failures spread through agent networks and why circuit breakers miss them.

Persistent agent memory requires built-in self-correction to stay reliable and safe.

MCP handles agent-to-tool access while A2A solves agent-to-agent coordination.

Agent credentials must stay separate so audit trails show which agent did what.

Structure matters more than model choice when frontier AI capabilities converge.

Stateless runtimes need external stores to persist agent memory across calls.

Preserve workflow progress across LLM timeouts by storing state durably instead of in memory.

Why standard containers fail when AI agents write their own code at runtime.

Splitting planning from execution trades speed for auditability.