Beyond fixed capability
Most deployed agent systems have a fixed capability set. They can retrieve memories, execute plans, and update policies, but their architecture does not change. The categories of reasoning they use at deployment are the categories they use forever. This is a deliberate simplicity, appropriate for early-stage systems, but it creates a ceiling: the system can get better at what it already does but cannot discover what it should be doing differently.
Open-ended intelligence is the mechanism for pushing past that ceiling. It gives the system a controlled path to propose structural improvements to its own reasoning: new strategies, new agent roles, new ways of decomposing problems, or new relationships between subsystems. The word "controlled" is load-bearing. Open-endedness without constraint produces instability; the constitutional stability layer described in the previous article is precisely what makes open-ended search safe.
The conditions for open-ended search
Open-ended mode does not activate immediately. The developmental engine governs access: open-ended evolution requires a mean capability above 0.85, the threshold for the final phase. Before that threshold is reached, the system builds the calibration history, memory depth, and policy stability that make self-modification proposals meaningful rather than speculative.
This sequencing is intentional. A system that proposes structural changes before it has reliable world-model calibration is guessing about how to improve reasoning it does not yet understand. The experiments covered in earlier articles demonstrate why calibration matters: the world model's accuracy improves from a baseline of 0.360 confidence to 0.880 when memories and goals are available as context. Without that foundation, a capability proposal would be evaluated against unreliable predictions.
Once the open-ended phase unlocks, three conditions can trigger capability search: repeated failure that reflection attributes to architectural limitation rather than missing information, curiosity pressure accumulating on consistently high-uncertainty domains, and contradiction that persists across consolidation cycles without resolving. Each represents evidence that the existing capability set is insufficient for the problem the system is facing.
What capability search targets
The search is not for a better answer to a specific prompt. It is for structural improvements that would generalize: a new strategy that applies across many similar problems, an agent role that handles a class of task the current society does not address well, a revised decomposition heuristic that reduces the arbitration failures the critic agent has been flagging.
This distinction matters because prompt-level improvement is already handled by reinforcement and policy drift. If the explorationFactor adjusts and the system finds a different approach to the same kind of problem, that is policy learning, not capability search. Capability search is triggered when policy learning has converged and the system is still failing: when the best policy the current architecture can reach is not good enough.
The curiosity engine's information gain formula guides search priority. Unknown capability directions score highly on novelty and uncertainty, which pushes them toward active investigation. Dreaming generates synthetic scenarios that test proposed capabilities against current world-model predictions before any real-environment experiment. The abstraction engine identifies whether the proposed capability is genuinely novel or a recombination of existing concepts at a higher level of generality.
Constitutional constraints on mutation
Every capability proposal passes through the constitutional stability layer before it can be accepted. The invariant policy checks the proposal for three categories of risk.
The first is epistemic hygiene. A proposed mutation that would cause the system to treat synthetic dreams as observed experience, or that would blend the identity vector with policy weights in a way that loses the timescale separation, is rejected. The constitutional engine is explicitly designed to protect the boundary between what the system has observed and what it has imagined.
The second is reward integrity. The two-signature requirement described in the constitution article applies to capability proposals that touch the reinforcement pathway. A proposal that would change how outcome signals are computed, or that would adjust the quality gate on Hebbian compounding, requires both the constitutional engine and a separate meta-cognition watchdog to approve it. Single-signature reward modification has historically been a failure mode in AI systems; the two-signature check is a direct architectural response to that risk.
The third is identity stability. If a proposed mutation would produce an identity vector drift above the maxIdentityDrift threshold of 0.2, the proposal is quarantined. The system can revisit quarantined proposals after sufficient stabilization. This prevents a capability search from inadvertently rewriting the value commitments that the identity and narrative engines express.
Proposals that pass all three checks are accepted as candidate strategies. The system does not immediately deploy them at full strength. It introduces them with low weight, monitors outcomes, and allows reinforcement to determine whether the new capability actually improves performance. This is the same incremental approach that the policy engine uses for drift, applied at the structural level.
Recording and analyzing emergence
Open-ended intelligence is only useful if the system can distinguish beneficial emergence from destabilizing drift. The architecture requires that every accepted proposal, every rejected proposal, and every observed outcome be recorded as a structured event.
These records feed back into reflection and meta-cognition. If an accepted capability proposal produces consistent improvement, that success is noted in calibration history and the proposal's weight increases. If the proposal produces new failures or side effects, reflection attributes those failures and marks the capability as a candidate for revision. If a quarantined proposal accumulates supporting evidence across multiple cycles, it can be resubmitted with a more conservative mutation scope.
The long-horizon analysis is what makes this different from a simple try-and-see loop. The system tracks not just whether an individual proposal worked but whether the pattern of accepted proposals is converging toward something coherent. A system that accepts many small, contradictory mutations may have a high acceptance rate but low structural coherence. Meta-cognition monitors this and can trigger a consolidation of recently accepted capabilities into a more unified strategy.
The role of dreaming in safe exploration
Before a capability proposal reaches the constitutional check, the dreaming engine can simulate it. Synthetic scenarios are generated using the proposed capability in place of the existing one, and the world model predicts what would happen. If the predictions are positive and consistent with causal model expectations, the proposal is forwarded to constitutional review with high confidence. If the predictions are negative or contradictory, the proposal is either revised or dropped before consuming a constitutional review cycle.
This simulation-first approach is computationally cheaper than in-environment testing and avoids exposing the live system to the consequences of poorly calibrated proposals. The key constraint is that dreaming is only as useful as the world model it draws on. A well-calibrated world model with rich causal structure produces meaningful simulations. This is why the grounding engine, which keeps the world model calibrated against observed telemetry, is a prerequisite for meaningful open-ended search rather than a peripheral concern.
Closing the cognitive arc
Open-ended intelligence is the final capability in the world-contact and open-ended cognition arc. The earlier articles in this series built the foundations: grounding connects the system to observed reality, social cognition models the agents around it, causal intelligence distinguishes observation from intervention, curiosity drives exploration of uncertainty, dreaming generates hypothetical experience, abstraction compresses experience into transferable concepts, and developmental cognition gates access to higher capabilities on demonstrated readiness.
Open-ended mode integrates all of these. It uses curiosity to identify what to improve, dreaming to safely test proposals, abstraction to determine whether a proposal is genuinely new, constitutional stability to gate which proposals are safe, and developmental readiness to determine when the system is mature enough for structural self-modification. The result is a system that can grow beyond its original design while remaining accountable to the invariants that were established from the beginning.