From Programmable Observation to Agentic Operations

Today’s question: if the human is not the consumer of all this very detailed fine-grained information, how do you synthesize state and decide what to do?

Today’s arc: the data plane became programmable; observation moved into it; ML decision-makers entered specific loops; a credibility crisis surfaced a gap the research frontier is still trying to close.

That paints a clean linear evolution with the programmable-data-plane shift as the prime mover. The actual story: parallel tracks running for decades, each chasing its own question, eventually fusing on a new substrate.

Three pressures shaped the trade for 30+ years:

L15 leaned on the human side. Cost is co-equal. Both pressures get relaxed across the rest of the lecture.

None of these needed programmable silicon. They predate it by decades. Each was answering a different question.

The question RMT’s designers were asking was structural, not measurement-specific:

If we expose the pipeline architecture as a target for software, what new design space opens up?

Sketches-in-switches and per-hop measurement came as applications later. An enabling shift, not the prime mover.

Deployment honesty: Tofino runs in production at Google, Meta, Microsoft datacenters. Enterprise switches, ISP edge routers, and consumer hardware mostly run fixed-function silicon. The work that follows matters intellectually; its operational footprint sits in a handful of organizations.

# Sonata: detect DNS amplification victims
victims = (packets
    .filter(lambda p: p.udp_sport == 53)
    .map(lambda p: (p.dst_ip, p.src_ip))
    .distinct()
    .reduce(keys=["dst_ip"], op=sum)
    .filter(lambda k, v: v > threshold))

Naive: millions of tuples/sec to server. Sonata’s compiler runs filter + distinct + most of the count on the switch — using Track 3 sketches — and ships only heavy hitters. ~10⁴× reduction in tuples-to-server. Switch resource budget and operator cognitive budget in the same ILP.

Sketches are exactly the right data structure for that distribution. A small fixed-size table gives bounded error on heavy-hitter queries precisely because heavy tails concentrate the mass in a few places.

The intellectual lag was ~20 years. The sketches existed first; the substrate let them be deployed.

Both remain the structural constraints the rest of the lecture is trying to relax further.

Geoff Huston (APNIC, 2024) called this the vanishing network. Now: who consumes the richer data?

Pattern across all three: narrow scope, defined fault model / decision space, ML beats hand-tuned heuristic on benchmarks. By 2021, every operator-facing loop above had a credible learned alternative.

A human still composes the four fragments into a single causal chain. Each learned component was trained on one fragment, not on the composition.

No single learned system assembles the cross-layer causal chain the way an experienced operator does.

That gap is what the next act turns into a structural question.

The crisis is structural, not anti-ML. Classifiers, anomaly detectors, NeuroPlan, AuTO, 007 — all worked on their benchmarks. The crisis: once the loop compresses past human reaction, you have to trust the machine; the old trust mechanisms (rate-limit, page someone, roll back) transfer poorly to learned components.

It forces four first-principles questions the engineering stack leaves open:

1. What is the right belief representation when the consumer is machine, not human?
2. What is the right decision-maker when no single program controls the whole loop?
3. What is the right verification regime over lossy learned components?
4. What is the right unit of attribution in a network of networks?

The four research fronts (Act 5) each take a swing at one sub-question. All four are in motion. The four answers don’t compose with each other yet.

Intended pipeline: A’s foundation model → B’s agent → C’s safety gate → D’s expert composition.

In theory. In practice, the layers lack interfaces:

• A safety gate (C) lacks a way to read a foundation-model embedding (A).
• An agent population (B) lacks a protocol for trusting an expert from a different stakeholder (D).

The composition problem is the real open frontier — downstream of all four. If you want to do research here, the open questions live at the joins, not at “another foundation model” or “another agent benchmark.”

The arc of machine consumers replacing humans in inner, then outer, then NOC-level loops is real. But the arc has a stopping point — and it’s not technical.

The outermost loop — intent, ethics, accountability, regulatory constraints — stays human. The stop comes from accountability, not capability. A campus IT department does not get to decide that a community should be deprioritized on the residential VLAN. That decision belongs to the people the community comprises, to the operators who answer to them, to the regulators who answer to all of them.

The operator’s role retreats outward, not away. The operator becomes the intent specifier and accountability gatekeeper. Agents handle the inner loops. Humans handle the loop that asks what the inner loops should optimize for, who they should serve, and who answers for them when they fail.

Week 10 is project work — extended office hours, bring your stuck moments. Thank you for the quarter. Good luck with your projects.