Most memory systems decide on the write path. An event happens, a model extracts what seems important, and the rest is discarded. That keeps the store clean. It also asks the write path to know, at ingest, what will matter in three months. A detail that looks irrelevant in a support chat may explain an escalation later. Once the raw event is gone, better retrieval cannot recover it.

The better pattern is the old data-lakehouse move: preserve what happened, then build the useful representations later. Summaries, embeddings, graphs, and memories become replaceable views over an immutable history. The same conversation can yield one set of facts today and another next year. Better models can re-index old history. New agents can use it in ways nobody planned at capture time.

A 4B coding agent, post-trained only with RL on synthetic software tasks, reaches 61.5% on SWE-bench Verified and beats models many times larger. No distillation from a frontier teacher. The missing ingredient is not a bigger model. It is tasks calibrated to what the current checkpoint can almost, but not quite, solve.

The pipeline estimates how often the agent already resolves a candidate task, then rewrites the ones that are too easy or too hard until they sit in a learnable band. As the policy improves, the curriculum moves: from heavily scaffolded bug fixes toward shorter statements and more diverse change types. A simple tool harness (read, write, edit, glob, bash) lifts the same 4B base far more than it lifts large models. Failures remain mostly reasoning, not tool use. Effective reward hacking is held at zero by the harness, not by hoping the agent is honest.

Recursive self-improvement is usually told as a spell: the model tweaks itself, gets smarter, tweaks itself again. Control theory treats that as a feedback loop with a maximum safe speed, set by how fast you can generate trustworthy evidence that a change was actually an improvement.

A loop stays stable when the product of its gains stays below one. Gain here is how large a modification the system makes per unit of evidence. The entire appeal of self-improvement is high gain. Adaptive control already learned the hard version of this: proofs assumed a plant, a mild violation of the assumptions sent the system haywire. Unstable adaptive systems often do not blow up immediately. They burst: run beautifully, then fail without warning.

If agents work with other agents, they need a standard way to talk. That was the case for an explicit agent-to-agent protocol. Adoption is real. Meanwhile the rest of the stack has reduced how often agents need to appear to one another as independently addressable peers.

A remote agent can look like a capability. Inference stays behind the interface. Subagents live inside one runtime and die with the task. A supervisor presents one identity while specialists stay internal. Software factories coordinate the work, not a mesh of persistent agent relationships. Workflow engines treat agents as steps in a graph they already own. Handoffs can be artifacts in durable state rather than messages. Queues and event buses already do this at scale.

You start in fog, on a dark map, with an objective and no knowledge of which direction is uphill. The craft has always been a loop: hypothesize, move, observe, update. Hill climbing is one part of that loop. It can take you to a local peak, or up the wrong mountain if the proxy is wrong. One company optimized registered users. Another optimized people who actually came back.

AI changes the economics of the loop, not the geometry of the terrain. Agents can analyze, propose, ship, and interpret. You can now climb the wrong hill much faster. A rigid roadmap assumes the order of tests matches the size of the opportunities. Bandits waste less traffic on losers. A better strategy spends a small slice of budget mapping lever sensitivity first: poke the map, then climb where it makes a difference.

Most companies will use AI to clear a faster ticket queue. In-house creative teams already function as service desks: hundreds of reasonable requests that reproduce decisions already made. When the queue is too slow, everyone else ships around it. The brand dies by papercuts. If the structure stays the same, AI only makes the same bad system move faster.

Guidelines were built for scarce production. They packaged judgment into a manual and handed it to people with different expertise. A guideline can describe the brand. It cannot participate in the work. Brand as software can: it carries beliefs, product truths, audience, voice, and the line where a human must step in into the artifact itself.

If machines take most economically necessary labour, relationships become the remaining source of structure and status. That is not automatically a kinder world. Pre-industrial life was relational: gossip, precedence, peer-group enforcement, and belonging that came with exclusion.

Look at people who already lived that way. Versailles courtiers. Upper-class volunteers on nonprofit boards. Working-class urban villagers whose jobs were precarious and whose real life was the peer group. Across class, the pattern is similar. Meaning came from each other. The world was small. Deviance was policed. Usefulness was often an illusion that kept the circle intact.

The usual case for health spending is that it raises productivity among people already in the measured economy. For a country trying to bring hundreds of millions more into productive life, that is too narrow. Illness changes household risk-taking: who migrates, who stays in school, who starts a business, who can leave unpaid care work. Health is infrastructure for participation, not a welfare add-on.

The deeper error is conflating money with systems. Financing builds facilities and fills posts. It does not create continuity, responsibility, or relationships across levels of care. Many states already spend enough, given cost advantages and labour slack. The scarce inputs are political capital (who has authority), regulatory capital (how incentives are aligned), and organizational capital (coordination, learning, accountability). Primary care fails without defined populations, risk stratification, and journeys that hold together. Digital tools do not substitute for that architecture.

Language models spend most of their attention on a small fraction of context, yet they still read the entire KV cache to find it. Ask about a detail in a long conversation and every token of the reply may require a full scan. Most sparse methods still pay an O(N) tax just to decide what matters.

Declarative Attention takes the opposite bet. The model already knows which parts are relevant. It declares the scope inside its chain of thought, in three modes. Global: look at everything to navigate. Focus: attend to a specific region. Local: use only recent output to synthesize. The inference engine parses those declarations like tool calls and skips most of the cache read.

The bottleneck on consequential decisions is not analysis. It is not knowing how people will actually behave before you spend. Traditional research is slow, expensive, and often measures what people say rather than what they do. Synthetic populations invert that. Query a modeled public overnight, then spend on the human study only if the stakes require it.

Two camps are forming. One interviews real people, then enriches and re-queries them. The other rebuilds populations from census, traces, and panels, arguing that interviews import the same bias simulation is meant to escape. Buyers are more pragmatic than philosophical. Accuracy, traceability, and whether the output can survive procurement are the first gates.