# Research

## Super Intelligence Lab

Our research agenda: six open problems that stand between today’s collections of capable agents and an organization that learns. For each we state the problem formally, say why current methods fall short, and describe the approach we are pursuing and how we will know it works.

Canonical: https://super-intelligence.inc/research

## 0 · Problem formulation

> **In plain words.** A company running many AI agents is a team without a manager. We study the manager: a layer that decides what each agent knows, what it may do, how its results are judged, and what the whole team remembers. Our measure of success is how fast the organization, as a whole, gets better at deciding.

We model an agentic organization as a decentralized, partially observable decision process with human principals:

$$
\mathcal{M} = \langle I, H, S, \{A_i\}, T, \{\Omega_i\}, O, U, C, B \rangle
$$

where $I$ is a set of AI agents, $H$ a set of human principals, $S$ the state of the business, $A_i$ and $\Omega_i$ each agent’s actions and observations, $T$ and $O$ the transition and observation functions, $U$ the organization’s utility, $C$ a cost vector (tokens and human attention) and $B$ its budgets. Three properties make the problem hard in practice. $U$ is not given: it is latent in past decisions and the people who made them. Outcomes are delayed, jointly produced and confounded. And exact decentralized control is intractable even in the idealized case: the finite-horizon Dec-POMDP is NEXP-complete [1].

_Figure 1 · The management policy μ sits between human principals and agents. It controls what agents know and may do, receives their outcomes, and reads and writes the decision record._

We therefore do not try to solve $\mathcal{M}$. We add structure. A **management policy** $\mu$ sits between $H$ and $I$ and controls four things: what each agent knows (context), what it is trying to achieve and is allowed to do (objectives and authority), how results are attributed (outcome), and what the organization retains (learning). The object we optimize is the learning curve of the whole organization.

$$
\mathrm{ROI}(n) = \frac{J(\pi_{n+\Delta}) - J(\pi_n)}{\Delta}
\quad \text{subject to} \quad
\mathbb{E}[C_{\text{tok}}] \le B_{\text{tok}},\;\;
\mathbb{E}[C_{\text{att}}] \le B_{\text{att}}
$$

Here $J(\pi_n)$ is the expected organizational utility of the joint policy after $n$ recorded decisions. **Recursive Organizational Intelligence** is the regime in which $\mathrm{ROI}(n)$ stays positive, improvement is non-regressing with high probability, and the components of $\mu$ are themselves trained on the decision record, so that improving the manager raises the rate of improvement. The six problems below are the conditions for reaching that regime.

_Figure 2 · Schematic. Organizational utility against recorded decisions. A model upgrade lifts both curves; the gap that remains is organizational learning, and its slope is ROI(n)._

### System view

In system terms, $\mu$ is built as two fabrics over one shared record. **Decision Intelligence** holds what the organization knows about deciding: codified domain knowledge (B1), the organization’s own decision system (B2), and learning from outcomes and, within privacy boundaries, across customers (B3). **Agent Harnessing** turns decisions into coordinated work: multi-agent execution (A1), coordination with people (A2) and bounded initiative (A3). Every step from context to review is written to a **Decision Ledger**, and nothing learned from it changes the system without passing a gate. Sections 1 and 3 build the decision fabric; Sections 2 and 6 build coordination among agents and with people; Section 4 fills the ledger’s outcome and review entries; Section 5 is the gated loop.

_Figure 3 · System overview. Decision Intelligence, the brain, decides and learns. Agent Harnessing, the nervous system and hands, coordinates and executes. Both write to one Decision Ledger, and lessons return to the decision fabric only through a gated loop: observe, propose, test, approve, promote._

## 1 · Ontology: an executable model of the organization

> **In plain words.** Before agents can work for a company, they have to mean the same thing the company means by its own words. We turn each business term into something a machine can check.

**Problem.** Recover and maintain a typed model of an organization’s concepts (goals, metrics, entities, policies) such that every agent action and every recorded decision refers to shared, checkable semantics.

**Why current methods fall short.** Glossaries in prompts are unversioned and unenforced. Embedding similarity recovers relatedness, not identity: “qualified lead” and “marketing-qualified lead” are close in vector space and different in the business. Hand-built schemas are correct on the day they ship and drift silently afterwards.

**Approach.**
- **Decision-contract-driven minimality.** We derive the ontology from decision contracts: the inputs, admissible actions, constraints and success measure of one concrete decision. A type exists only if some decision needs it. This keeps the ontology small enough to be falsified.
- **Triple grounding.** Each concept $c$ carries an intensional definition (language), an extensional definition (an executable predicate over the systems of record) and a set of human-ratified exemplars. Disagreement among the three is a measurable error signal, not a matter of opinion.
- **Induction and alignment.** Candidate types and relations are proposed from schemas, documents, conversations and agent traces, then reconciled with probabilistic record linkage [2] and ontology matching [3]. Where two teams define a concept differently, the system creates an explicit _contested concept_ with a named human owner. Conflicts are never merged silently.
- **Bitemporal versioning.** Every assertion carries valid time and transaction time [4], so any past decision can be replayed against the ontology as it was believed at the moment of decision.
- **Drift detection.** We monitor the extension of each concept and test for distribution shift in how agents and people apply it. A drifting concept is sent back to its owner for re-ratification before agents may rely on it.

_Figure 4 · (a) Each concept is grounded three ways, and disagreement between any two is a measurable error. (b) Bitemporal versioning: a correction made in March changes what is believed about February without erasing what was believed before._

**How we measure progress.** Agreement between agent usage and concept owners (Cohen’s $\kappa$); precision and recall of executable definitions against exemplars; time to detect an injected definitional change; replay fidelity of historical decisions.

**Open questions.** How to represent tacit norms that no one has written down; how definitions compose; how small an ontology can be while still supporting transfer across decisions.

## 2 · Context: budgeted information routing

> **In plain words.** An agent that knows too little drifts; an agent told everything loses the thread. We send each agent, and each person, only what would change their next decision.

**Problem.** Each agent $i$ has a finite context budget $B_i$; each person has a finite attention budget. Given a shared organizational state $Z$, decide which information reaches whom, and when. Inter-agent misalignment is one of the three root failure categories identified across more than 1,600 multi-agent execution traces [5].

**Formulation.** For agent $i$ facing decision $d$, choose a message $m_i \subseteq Z$ that maximizes the value of information [6] net of its cost:

$$
m_i^* = \arg\max_{m \subseteq Z} \; \mathbb{E}\big[\,U(a_i^*(m))\,\big] - \lambda\,|m|
\quad \text{s.t.} \quad |m| \le B_i
$$

which we approximate with an information-bottleneck objective [7]: compress $Z$ into $M_i$ while preserving information about the agent’s optimal action, $\min\; I(Z; M_i) - \beta\, I(M_i; A_i^*)$.

**Approach.**
- **Shared state as a typed event log.** Agents hold materialized views of an append-only log. Commutative concurrent updates merge with conflict-free replicated data type semantics [8]. Non-commutative conflicts, such as two agents committing incompatible plans, are surfaced to the manager as first-class objects.
- **Relevance learned from counterfactuals.** For each event $e$ and agent $i$ we estimate decision sensitivity by replaying decisions with the event ablated: $\Delta_i(e) = D_{\mathrm{KL}}\big(\pi_i(\cdot \mid c)\,\|\,\pi_i(\cdot \mid c \setminus e)\big)$. A router is trained to predict $\Delta$ and forwards an event only when its predicted value exceeds its token cost.
- **Decision-preserving compression.** Histories are compressed under the constraint that the downstream decision distribution is unchanged within $\varepsilon$. Textual fidelity is not the objective.
- **Freshness contracts.** Facts are causally consistent. An agent may not commit an action whose preconditions rest on a view older than a per-decision bound $\tau$.
- **One router for agents and people.** The same mechanism decides what reaches which human. Attention is a budget in the objective, not an afterthought.

_Figure 5 · (a) Events are forwarded only to agents whose decision they would change, within each agent’s budget. (b) Decision quality against context size: sharing everything first helps, then hurts; routing and compression reach higher quality with fewer tokens._

**How we measure progress.** Inter-agent misalignment rate under the taxonomy of [5]; decision divergence under ablation; tokens per decision at fixed decision quality; number of actions committed on stale preconditions.

**Open questions.** Relevance is non-stationary; $\varepsilon$-preservation does not obviously compose across long chains of agents; shared state is an attack surface, so context poisoning needs its own defenses.

## 3 · Decisions: precedent-grounded judgment with calibrated abstention

> **In plain words.** A good employee decides the way the company has decided before, and asks when unsure. We build both habits: decisions grounded in past cases, and a mathematically calibrated way to say “I should ask”.

**Problem.** Produce decisions that are consistent with the organization’s own precedent, informed by what works across the industry, and calibrated: the system must know when not to decide.

**Why current methods fall short.** A general model supplies general judgment. Retrieval over documents returns text, not precedent. And the self-reported confidence of language models is not a reliable basis for delegating authority.

**Approach.**
- **The decision record as a first-class object.** $\langle$ situation $x$, evidence $E$, action $a$, alternatives considered, stated rationale $r$, authority, policy version, outcome $y$, review $\rangle$. The rationale is stored as a claim, never as a causal edge; causal edges are produced only by outcome analysis (Section 4).
- **Retrieval as learned case-based reasoning** [9]. Two cases are similar if the same action was right for the same reasons. Retrieval returns both supporting and opposing precedents.
- **Defeasible conflict resolution.** When precedents disagree, an explicit ordering applies (specificity, then authority, then recency). Every override is logged as new precedent.
- **Hierarchical judgment.** Company-level parameters are partially pooled toward an industry-level prior, $\theta_c \sim \mathcal{N}(\theta_{\text{ind}}, \Sigma)$ [10]. Cross-organization learning uses only aggregated, de-identified statistics.
- **Calibrated abstention.** Conformal prediction [11, 12] over the admissible action set yields a set $\mathcal{C}(x)$ with coverage at least $1 - \alpha$. A singleton set is acted on, within authority. A larger set is escalated together with the precedents that separate the candidates.
- **Consistency audits.** Counterfactual twin tests perturb a situation in decision-irrelevant ways and require invariance, then in decision-relevant ways and require the documented change.

_Figure 6 · (a) Judgment parameters start at a wide industry prior and move to a narrower company posterior as decisions accumulate. (b) A conformal set with one admissible action is acted on; a set with several is escalated._

**How we measure progress.** Agreement with held-out human decisions; invariance under irrelevant perturbation; empirical coverage and mean size of $\mathcal{C}(x)$; escalation precision, defined as the fraction of escalations in which the human changed the default.

**Open questions.** Exchangeability fails under drift, which calls for adaptive conformal methods; some precedents were authoritative and wrong; value judgments and factual judgments may need different machinery.

## 4 · Outcome: attribution under delay, noise and confounding

> **In plain words.** Results come in weeks later, mixed up with everything else that happened. We work out what each decision and each agent actually contributed, and what it cost.

**Problem.** Estimate the value of a decision policy, and each agent’s contribution to it, when outcomes are delayed, jointly produced and confounded, and when any metric that is optimized will be gamed [13].

**Approach.**
- **Off-policy estimation from the decision log.** The manager logs its own action propensities at decision time, which permits inverse-propensity and doubly robust estimators [14]:

  $$
  \hat{V}_{\mathrm{DR}}(\pi) = \frac{1}{n} \sum_{t} \Big[ \hat{q}(x_t, \pi) + \frac{\pi(a_t \mid x_t)}{\mu(a_t \mid x_t)} \big( y_t - \hat{q}(x_t, a_t) \big) \Big]
  $$

  This requires a small amount of deliberate exploration. We treat the exploration budget as an explicit parameter that a human approves.
- **Identification without randomization.** Where experiments are not permitted we use difference-in-differences and synthetic control [15]. Every causal conclusion is stored with its identification strategy and assumptions [16].
- **Delayed outcomes.** A surrogate index [17] combines short-term signals into a predictor of the long-term outcome, re-validated against realized outcomes.
- **Multi-agent credit assignment.** An agent’s contribution is estimated against a counterfactual baseline [18]; where agents interact strongly we use sampled Shapley attribution [19].
- **Resistance to metric gaming.** Metric ensembles with held-out guardrail metrics that agents never observe, periodic adversarial audits, and human review sampled in proportion to an anomaly score.
- **Cost as a constraint.** Token and attention costs enter a constrained MDP [20] solved by Lagrangian relaxation. The multipliers are shadow prices: what one more token, or one more minute of an expert, is currently worth. We show them to people directly.

_Figure 7 · (a) Short-term signals are combined into a surrogate that predicts the long-term outcome. (b) The utility–cost frontier under a budget; the slope at the budget is the shadow price of one more token or one more expert minute._

**How we measure progress.** Bias and variance of estimators against randomized ground truth where it exists; validity of surrogates over time; stability of attributions under resampling; cost per unit of outcome.

**Open questions.** Small-$n$ regimes, since an organization makes few large decisions; interference between units; persistent non-stationarity.

## 5 · Self-improvement: recursive learning without regression

> **In plain words.** When one agent finds something that works, every agent should have it by morning, and nothing that already worked should break. A lesson has to pass three tests before it spreads.

**Problem.** Turn experience into shared capability that improves every agent, with a guarantee that new learning does not break existing competence.

**Why current methods fall short.** Continual learning in the weights of a single model is costly, opaque and prone to catastrophic forgetting [21]. Appending to memory grows without bound and accumulates contradictions. Neither transfers what one agent learned to another.

**Approach.**
- **Three loci of learning, ordered by reversibility.** L0, knowledge: ontology, precedents, playbooks. L1, harness: routing, tool policies, workflows. L2, parameters: small vertical models adapted on decision records. A gain is attributed to the most reversible locus that explains it.
- **Experience distillation.** Successful episodes are clustered and abstracted into parameterized playbooks. A playbook is a hypothesis.
- **A promotion gate.** A hypothesis enters the shared library only if its uplift is significant under anytime-valid sequential inference [22], it transfers to other agents’ task distributions in shadow mode, and it passes a non-regression replay suite built from historical decisions.
- **Exploration at the level of the organization.** Variants are allocated across agents by posterior sampling [23]; the organization, not the single agent, is the unit of exploration.
- **Diversity preservation.** A quality-diversity archive [24] keeps behaviorally distinct strategies alive.
- **Forgetting by design.** Memories carry utility estimates and decay. Contradicted items are retired with tombstones for audit. Every state is versioned and can be rolled back.
- **The recursion.** The router of Section 2, the retrieval metric of Section 3 and the estimators of Section 4 are trained on the same decision log. Improving them raises $\mathrm{ROI}(n)$. Changes at L1 and L2 pass the same promotion gate and require human sign-off.

_Figure 8 · (a) A lesson must pass three gates before it reaches every agent. (b) Learning is applied at the most reversible locus that explains the gain; less reversible loci face a stricter gate._

**How we measure progress.** Slope of the organizational learning curve $\mathrm{ROI}(n)$; transfer ratio, the gain on non-originating agents divided by the gain on the originating agent; regression rate per promotion; time from first success to organization-wide default.

**Open questions.** Formal bounds on safe self-modification; early detection of negative transfer; how to credit exploration that pays off only for other agents.

## 6 · Expert in the loop: human attention as the scarce resource

> **In plain words.** An expert’s minutes are the most expensive input in the system. We ask only when it matters, learn the most from every answer, and let the agent earn independence one kind of action at a time.

**Problem.** Decide when a person must decide, which person, and with what evidence; learn as much as possible from each intervention; and let autonomy be earned and revoked.

**Approach.**
- **Learning to defer.** The deferral policy is trained jointly with the decision policy to minimize system loss, including the expert’s own error and the cost of their attention [25, 26].
- **Risk-budgeted graduated autonomy.** Each action class $k$ has an authority tier: act, ask or never. Promotion from ask to act requires that the lower confidence bound on supervised success over $n$ decisions exceeds a threshold $\theta_k$, and that the action is reversible. Demotion on violation is immediate. The never tier is a hard constraint outside the learner.
- **Active querying.** Experts are asked where expected information gain per minute is highest: $\mathrm{score}(q) = \mathrm{EIG}(q) / \mathrm{cost}(q)$.
- **Corrections as the highest-value data.** An override is a labelled counterexample with a rationale, written to the precedent store with the expert’s authority weight.
- **Reporting as faithful compression.** A brief $b$ of a trace $\tau$ is acceptable when the expert’s decision given $b$ equals their decision given $\tau$. We call this decision equivalence.
- **Guarding against automation bias.** Occasional challenge cases, opposing precedents shown by default, and tracking of the rate of unexamined approvals.

_Figure 9 · (a) Authority is earned per action class and revoked at once on violation; the never tier sits outside the learner. (b) The share of decisions escalated to a person falls as classes are promoted and rises after a violation._

**How we measure progress.** Joint human-and-AI decision quality against either alone; attention minutes per unit of outcome; escalation precision and recall; decision-equivalence rate of briefs on audited samples; time to autonomy per action class.

**Open questions.** What to do when experts disagree; how to detect that an expert’s own judgment has drifted; legitimacy, meaning who is entitled to overrule whom.

## 7 · What would prove us wrong

A research agenda should be falsifiable. We would revise ours if any of the following held on real organizational data:
- The organizational learning curve is flat: $\mathrm{ROI}(n)$ is statistically indistinguishable from zero after controlling for improvements in the underlying models.
- Transfer ratio stays near zero, meaning lessons do not move between agents and the organization is no more than the sum of its parts.
- Calibrated abstention escalates so often that human attention, not model capability, becomes the binding constraint with no downward trend.
- Stronger foundation models alone close the gap in consistency with an organization’s own precedent, without any decision record.

## References

1. Bernstein, Givan, Immerman, Zilberstein. The Complexity of Decentralized Control of Markov Decision Processes. _Mathematics of Operations Research_, 2002.
2. Fellegi, Sunter. A Theory for Record Linkage. _Journal of the American Statistical Association_, 1969.
3. Euzenat, Shvaiko. _Ontology Matching_. Springer, 2007.
4. Snodgrass. _Developing Time-Oriented Database Applications in SQL_. Morgan Kaufmann, 1999.
5. Cemri, Pan, Yang, et al. Why Do Multi-Agent LLM Systems Fail? arXiv:2503.13657, 2025.
6. Howard. Information Value Theory. _IEEE Transactions on Systems Science and Cybernetics_, 1966.
7. Tishby, Pereira, Bialek. The Information Bottleneck Method. _Allerton Conference_, 1999.
8. Shapiro, Preguiça, Baquero, Zawirski. Conflict-Free Replicated Data Types. _SSS_, 2011.
9. Aamodt, Plaza. Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches. _AI Communications_, 1994.
10. Gelman, Hill. _Data Analysis Using Regression and Multilevel/Hierarchical Models_. Cambridge University Press, 2007.
11. Vovk, Gammerman, Shafer. _Algorithmic Learning in a Random World_. Springer, 2005.
12. Angelopoulos, Bates. A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification. arXiv:2107.07511, 2021.
13. Manheim, Garrabrant. Categorizing Variants of Goodhart’s Law. arXiv:1803.04585, 2018.
14. Dudík, Langford, Li. Doubly Robust Policy Evaluation and Learning. _ICML_, 2011.
15. Abadie, Diamond, Hainmueller. Synthetic Control Methods for Comparative Case Studies. _Journal of the American Statistical Association_, 2010.
16. Pearl. _Causality_, 2nd ed. Cambridge University Press, 2009.
17. Athey, Chetty, Imbens, Kang. The Surrogate Index: Combining Short-Term Proxies to Estimate Long-Term Treatment Effects More Rapidly and Precisely. NBER Working Paper 26463, 2019.
18. Foerster, Farquhar, Afouras, Nardelli, Whiteson. Counterfactual Multi-Agent Policy Gradients. _AAAI_, 2018.
19. Shapley. A Value for n-Person Games. _Contributions to the Theory of Games II_, 1953.
20. Altman. _Constrained Markov Decision Processes_. Chapman & Hall, 1999.
21. Kirkpatrick, Pascanu, Rabinowitz, et al. Overcoming Catastrophic Forgetting in Neural Networks. _PNAS_, 2017.
22. Johari, Koomen, Pekelis, Walsh. Always Valid Inference: Continuous Monitoring of A/B Tests. _Operations Research_, 2022.
23. Thompson. On the Likelihood that One Unknown Probability Exceeds Another in View of the Evidence of Two Samples. _Biometrika_, 1933.
24. Pugh, Soros, Stanley. Quality Diversity: A New Frontier for Evolutionary Computation. _Frontiers in Robotics and AI_, 2016.
25. Madras, Pitassi, Zemel. Predict Responsibly: Improving Fairness and Accuracy by Learning to Defer. _NeurIPS_, 2018.
26. Mozannar, Sontag. Consistent Estimators for Learning to Defer to an Expert. _ICML_, 2020.

## See the research at work.

We will show you the Management Agent on your own agents and decisions.