Reading 04 · Agents
Agentic AI — New Wine in an Old Bottle?
LLMs bring real gains in perception and adaptivity. But the hard engineering problems of early multi-agent research — coordination, knowledge, autonomy — remain stubbornly unsolved beneath the new vocabulary.
The thesis
Agentic AI is the second avatar of the agent-systems idea — modern deep learning applied to longstanding architectural problems, rather than a fundamental solution to them.
The argument, in five moves
Old problems beneath the new wine
Continuity, not rupture
Agentic AI's core challenges echo problems named in early-2000s distributed-agent research; the field applies contemporary learning to old architecture rather than resolving it.
Two layers, endlessly conflated
Building agentic systems demands mastery of both infrastructure — distributed systems, compute, storage, messaging — and a knowledge layer of semantics, decision-making and coordination. Most confusion comes from mixing the two.
The knowledge layer is the bottleneck
Distributed systems are well understood — CAP, replication, protocols. What resists us is decomposing expertise, handling adaptivity, coordinating collaboration and fusing symbolic reasoning with neural learning.
Autonomy is still low
Today's orchestration platforms show only basic coordination, comparable to earlier workflow systems. There are no standard autonomy metrics, and most decisions are pre-configured by engineers.
Deployment is expensive and risky
Production costs often exceed human-only alternatives; testing, validation, knowledge updates, failure recovery and security remain formidable and unresolved.
Symbol-weak systems, speaking no shared tongue — the “agent babel” that still awaits its PNG.
The Nyāya lens · शब्द
Nyāya counts śabda — reliable testimony through shared, interpretable language — among the valid means of knowledge. Agent babel is precisely a failure of śabda: without a common symbol system, agents cannot transmit knowledge that another can trust. The standard we lack is epistemic before it is technical.
Distilled from the original essay on engkraft.com.