From 2016 energy orchestration to 2026 agentic systems

Intelligence that respects cost and locality compounds. Intelligence that taxes everyone to “know everything” eventually serves almost no one.

The enduring constraint

In 2016 the observation was simple: smartphones are energy-constrained systems whose power behaviour is chaotic and multi-component. Continuous fine-grained measurement itself consumes scarce resources. The same physics still governs every modern AI agent stack.

Context windows, tool calls, memory retrieval, model escalations, and continuous sensing all cost energy, latency, tokens, and money. The more an agent tries to “know everything all the time,” the more expensive it becomes. The sustainable path is a learning loop in which the cost of acquiring the next unit of useful information falls while decision quality rises—so capability can widen without devouring the surplus that people actually need.

The original virtuous cycle

EEaaS argued that devices should not stream raw telemetry forever. They report selectively; the cloud learns which signals actually matter (Key Energy Indicators); devices progressively report less. Monitoring overhead shrinks while power-management decisions improve—getting better at measuring less while deciding better.

Mapping onto agent workflows

That principle maps cleanly onto today’s runtimes. A 2026 equivalent of Key Energy Indicators is a set of Key Decision Indicators (KDIs): the minimal observations—tool outcomes, user corrections, confidence signals, context hashes—that actually change future policy. Agents should start broad, then prune logging and memory writes as each signal class earns its keep.

On-device models handle routine work. Only when local confidence is low, or a KDI threshold is crossed, does the agent escalate to a larger cloud model or external tool. Memory writes, embeddings, and re-ranking are treated as first-class energy and token consumers: low-value memories compact or disappear; high-value patterns stay—and may be shared under privacy constraints without continuous centralisation of raw traces.

A modern loop for long-lived agents

An agent system optimised for decreasing learning cost moves through four phases—each one a deliberate act of making the next unit of understanding cheaper for everyone who depends on it:

1. Broad but cheap observationWide, low-resolution signals: tool success/failure, latency buckets, accept/reject, local confidence.
2. Importance learningOffline or low-priority analysis ranks signals by contribution to decision quality.
3. Policy tighteningLogging, memory write, and escalation rules focus on high-value signals; budgets shrink.
4. Continuous but sparse improvementEdge cases still get richer traces, but steady-state learning cost keeps falling while average decision quality rises.

Why the principle is more urgent now

Scale of cost: a poorly instrumented multi-agent workflow can burn orders of magnitude more energy and money than a 2016 smartphone profiler—waste that shows up as price, latency, and exclusion. Local-first expectation: people want capable agents that work offline; continuous cloud dependence is expensive and fragile. Compounding lifetime: agents accumulate memory and policy over weeks; without an explicit mechanism that reduces the cost of further learning, history itself becomes a liability.

Why this still holds in two to five years

The stack diagrams will age. On-device SLMs, remote giants, and the names we give indicators will shuffle. The constraint will not. Observation is never free—tokens, joules, latency, attention, and memory writes remain scarce in some combination. Cheaper inference moves the bottleneck; it does not abolish it. “Know everything always” is still an economic failure mode: at multi-agent, long-lived scale, unselective logging turns history into liability. That gets more true as agents persist for weeks and months—not less.

Client context stays first-class even if 99.99% of FLOPs run on Earth or in the Space. Answers still have to land on phones, laptops, IoT, intermittent links, and privacy boundaries. Providers that ignore those conditions waste capacity—or fail people.

Even the form of those clients can change completely. AI alone—or AI as a partner in invention—may produce a new class of hardware the way button phones gave way to touch screens, then to watches, earbuds, glasses, and whatever comes next. That does not retire this essay. New devices still live under energy, latency, privacy, and attention limits; they still need selective sensing; intelligence that lands on them still has to respect the conditions of the person holding (or wearing) them. The substrate reshuffles. The demand that learning get cheaper while decisions get better does not.

Closing

The durable contribution of the 2016 EEaaS work is not any particular logging API. It is the recognition that energy-aware—and today token- and latency-aware—systems must optimise the learning process itself.

When the cost of learning falls while the quality of decisions rises, intelligence becomes sustainable—and therefore able to serve the greatest number of people for the longest time. When the opposite occurs, every additional capability collapses under its own observational overhead.