The new set of problems for engineers to solve

gist

Katelyn Lesse, Anthropic's head of platform, argues that large multiagent systems create engineering problems only software engineers can solve. In an X Article she names four: infrastructure for stateful, bursty agent workloads; agent identity and authorization; security and isolation against compromised agents; and monitoring with kill switches. She pitches this as good news for engineers worried their skills no longer matter.

ideas

• Kubernetes fits agents poorly. Agents are isolated, stateful, single-instance workloads that idle unpredictably while waiting on tools, inference, other agents or humans; she points to new filesystems, sandboxing products and Google's Agent Substrate runtime as early responses.
• Authorization is the open identity problem. Authentication can borrow microservice workload identity with short-lived certs, but agents acting on a human's behalf raise consent questions for weeks-long runs, and agent-to-agent trust is even less solved.
• Multiagent systems widen the blast radius. Vulnerabilities are surfacing faster than patches; a compromised agent can pass injected instructions to peers or into shared memory, so agents should act in strongly isolated sandboxes and core software needs hardening.
• Monitoring needs tamper-proof traces and kill switches. She calls for tracing across agents that no agent can alter and kill switches that halt the whole system and cut its credentials; the hard part is telling persistence from malice.

quotes