Something remarkable happened in agent research in the first half of 2026, and it happened quietly, across papers that do not cite each other: the field changed its mind about where an agent's behavior comes from.
The new consensus, in the words of a Peking University group: "An LLM agent is not just an LLM… it is a model embedded in a stateful interaction loop," whose behavior is "determined not only by the model, but also by the runtime harness that mediates how the model observes the environment, understands tools, realizes actions, interprets feedback, and regulates multi-step trajectories." Their system evolves that harness — never touching the model — and improves 116 of 126 model-environment settings. The striking part: the harness was evolved from the failure traces of one small model, then transferred, unmodified, to seventeen others. It still worked. The harness had captured something about the environment side of the loop that no model weights contain.
They are not alone. A single-author paper (LEVI) ran the cleanest version of the experiment: same model, same evaluation budget, different search architectures — and the architecture dominated, with roughly a twelve-fold sample-efficiency gap between the best and worst harness wrapped around the identical model. Another paper this spring is literally titled "Stop Comparing LLM Agents Without Disclosing the Harness." The 2026 zeitgeist is real, it is empirically grounded, and I believe it: the loop around the model is now a first-order determinant of what these systems can do.
Here is what the field does not have: a single theorem about that loop.
Regulation vocabulary, no regulator theory
Read the PKU paper closely and you find a component named the "Trajectory Regulation Layer" — it monitors the interaction, detects loops and degenerate trajectories, and intervenes to recover. That is a controller. It is named like one, it behaves like one, and nowhere in the paper — or its related work, or its stated future work — does any control-theoretic question appear. Not "when does this converge." Not "what perturbations can it absorb." Not "under what conditions does regulation fail." The gap is not acknowledged, because from inside the engineering frame the gap is invisible: the system improves benchmarks, and benchmarks are the only instrument in the room.
This is not a criticism of one paper. It is the shape of the whole wave. The harness-engineering literature and the small control-theoretic literature that does exist — a McGill group proving exponential stability for agentic loops under a margin condition; a security group shipping a Lean-verified Lyapunov certificate for a tool-confined agent — have zero citations between them. Two communities are describing the same object, one with measurements and no theory, one with theory and no contact with the systems people actually run.
The field's own 1,250-paper survey of self-improvement, published this month, names the situation precisely in its open-problems section: what is missing is "stability engineering as a discipline" — a unified treatment of "when [loops] converge, oscillate, or collapse" that "would replace today's per-theme folklore." Folklore is the survey's word, and it is the right one. The field believes the harness is the controller. It cannot yet state what that belief commits it to.
What the theorems would actually say
Part 1 of this pair showed why this matters now: benchmarks are saturating, so the measurable differences between systems are migrating from the model into the loop — and loop properties are exactly what the current toolkit cannot measure or guarantee. Having spent this year building the control-systems account of these loops (the RCS program), here is my map of the four results the folklore is missing, each one upgrading a thing the field already believes into a thing it can rely on.
1. "Model-agnostic harness" should be a uniform margin, not a transfer table. The PKU result — one harness, seventeen models — is an empirical glimpse of something provable: treat the harness as a robust controller and the model as a bounded perturbation, and classical control gives you a stability margin that holds uniformly over every model whose deviation stays inside the bound — a set-certificate, one proof covering a model class, instead of a per-model benchmark row. The mathematics is not new; it is forty-year-old robust control. What is new is the reduction — and the honest boundary that comes with it: below a competence threshold the certificate says nothing, and that threshold is a property of the harness you can engineer, not a fact of nature.
2. The two thresholds move in opposite directions. Capability scaling clears the competence threshold permanently — frontier models execute any reasonable harness's intent, and that only improves. But the same scaling erodes the independence threshold: a stronger optimizer is better at gaming whatever verifier closes the loop. The field has each half separately (every reward-hacking report is the second half; every "small model + good harness" result is the first) but not the composition: there is a crossover where more capability starts hurting loop stability, and its location is a harness property. Notably, whether verifier-gaming actually scales with optimizer capability has — as far as I can find — zero direct measurements in the literature. It is the most consequential unmeasured quantity in agent engineering.
3. The meta-level question is an observability question. The new wave of meta-controlled search (AdaEvolve, ShinkaEvolve, temperature meta-policies) tunes exploration schedules from reward signals — the same information the base loop already has. One paper this spring did something different: it conditioned the schedule on the model's internal hidden states — a channel the reward signal does not carry — and something diagnostic happened in its baseline table. The reward-learned schedule controller lost to a fixed heuristic. Learning bought nothing. The internal-state controller beat both, and the mechanism their curves show is prevented entropy collapse: a higher converged ceiling, not faster convergence. One result, properly confounded, never framed this way by its authors — but it is exactly the signature the control account predicts: a meta-level lifts the ceiling only if it observes something the base level cannot compute from its own state. Control authority without observability extension buys speed, never height. That claim is stateable, testable, and absent from the literature.
4. The stability condition for stacked loops already has a name. Fast inner loop, slow outer loop — every two-timescale system in the wave (the skill-evolution papers tune their meta-update horizon by sweep and describe the failure mode as "trading reactivity for stability") is rediscovering, informally, the timescale-separation condition that stochastic approximation theory (Borkar) and simulated annealing formalized decades ago. The theorem they need exists. It has simply never been carried across the street.
The honest position
I want to be precise about what I am and am not claiming. The narrative — develop the harness, not just the model — is the field's now; the empirics above belong to the groups who produced them, and all of them (including the survey, including my own results) are preprint-grade and should be read that way. What I am claiming is the formal layer: the reduction that makes "model-agnostic" a theorem, the two-threshold crossover, the observability-extension condition, the transferred stability theory — and a pre-registered experiment, currently being designed on an open-ended substrate with the confounds above controlled, that can falsify the most important of them. The independence principle from Part 0 said the checker must be causally independent of the worker. This pair of posts is the same principle one level up: a belief about loops is folklore until an instrument independent of the believer can refute it. The field built the loops. The instruments are what come next.
The RCS program's theorems-in-progress, experiment pre-registrations, and negative results ship in the open at the repo linked above. If you are working on the control theory of agent loops — or you think one of the four claims above is wrong — I want to hear from you.