Part I

One score is not the road

“General” can describe breadth of tasks, depth of performance, autonomy of operation or transfer to unfamiliar settings. These properties need not rise together. The Levels of AGI framework explicitly separates performance, generality and autonomy, which already warns against treating one leaderboard as a scalar measure of the whole system.¹ A system may answer across many domains while needing a person to frame every task. Another may operate for hours inside one repository while failing outside its familiar tools.

Benchmarks are valuable precisely because they cut out a measurable question. GAIA combines reasoning, multimodal evidence, browsing and tool use in real-world questions.² OSWorld places multimodal agents in executable desktop environments; its original release found a large human-agent gap and attributed much of the agent failure to GUI grounding and operational knowledge.³ These results reveal missing competence. They do not produce a universal conversion from benchmark points to safe hours of autonomous work.

METR's task-completion time horizon offers a particularly useful measurement: task difficulty is expressed in the time a skilled human needs, and model success is fitted against that duration. The current official explanation is equally important. The metric is not the literal time an agent can run unattended, its tasks are concentrated in software engineering, machine learning and cybersecurity, and measurements beyond sixteen hours are currently described as unreliable for the task suite.⁴ A strong metric remains bounded by its task distribution, success definition and operating scaffold.

Thought experiment one: the hundred-step clerk

Imagine a clerk that performs every isolated action correctly 99 per cent of the time. Give it one action and it looks excellent. Give it one hundred necessary actions, with no checkpoint, retry or recovery. Under an illustrative independence assumption, the chance that every action succeeds is 0.99^100, about 36.6 per cent. Change only the horizon; the local skill remains fixed.

Real trajectories are neither identical nor independent. A bad early observation can corrupt many later steps, while a good checkpoint can prevent a local error from spreading. The simple calculation is therefore not a forecast. It exposes a structural point: long-horizon competence cannot be inferred by extending a short-task score. The system needs state reconstruction, progress tests, selective verification, recovery and a stopping rule.

Here, Gᵢ is the event that required gap i remains within tolerance; x is the task and environment; π is the configured policy and scaffold; and e is the evidence available at that point. The conditional form matters. If a grounding error changes the world state seen later, the world-model and planning events are evaluated under a different history.

The negative lesson is not that progress is illusory. Current time-horizon measurements show substantial gains on their evaluated task distribution.⁴ Work on adaptive test-time compute also finds that selective spending can outperform less discriminating inference on difficult, verifiable problems.⁵ The lesson is narrower: a curve in one slice becomes evidence for general agency only after it survives changes in environment, outcome definition, horizon, tool set, human context and consequence.

Part II

flowchart TD
  A["The relief convoy that solved<br/>the wrong world"] --> B["One score is not the road"]
  B --> C["Six gaps that must meet"]
  C --> D["Why the gaps couple"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#5B5CF6,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

Six gaps that must meet

The six gaps below are not six product features. Each names a causal burden that can be isolated through intervention. A system closes a gap only when changing the relevant condition changes measured behaviour in the predicted direction. A convincing demonstration should also include a negative control, because impressive behaviour may come from a shortcut that makes the proposed mechanism unnecessary.

1. Long-horizon reliability: preserve a task through change

A long task is not merely a short task with more tokens. It creates dependency. An early choice changes later options; a timeout can leave an unknown effect; a compacted history can omit why a branch was rejected; the environment may move while the system is reasoning. Long-horizon reliability is the ability to maintain a correct-enough task state, detect divergence and resume without duplicating or reviving invalid work.

Recent general-agent evaluation reinforces the distinction. General AgentBench reports that leading agents degrade when several skills and tools must operate in one unified setting. In that study, sequential scaling encountered a context ceiling, while parallel sampling encountered a verification gap.⁶ Those are architectural failure descriptions. More steps and more samples did not automatically create a better maintained episode.

The leading indicator is not transcript length. It is reliability conditional on horizon and injected disturbance. Hold task family and resources as constant as possible. Increase dependency depth, delay one tool, invalidate one assumption and interrupt the run after a committed effect. Measure recovery, duplicate effects, forgotten constraints, false completion and unresolved unknown outcomes. A system that succeeds only when nothing changes has learned a route, not robust execution.

2. World modelling: predict what intervention changes

A world model supports a claim of the form: if this action occurs in this state, these future states become more or less likely. Familiar continuation is not enough. The model must preserve objects, hidden variables, causal constraints and persistence when an intervention breaks the usual visual or textual pattern.

WorldPrediction tests high-level world modelling and procedural planning using counterfactual distractors. Its reported frontier-model results were much lower on ordered procedural planning than on simpler world-state discrimination, while human evaluators solved the benchmark in the study.⁷ DreamerV3 supplies a strong positive counterpoint: a learned world model with one configuration performed across more than 150 tasks and achieved a notable Minecraft result by imagining future trajectories.⁸ That is evidence that learned dynamics can support broad control inside specified environments. It does not establish an unrestricted model of the open world.

Thought experiment two: the swapped hinge

Place the same agent in two visually identical rooms. In both, a cabinet door has the same handle and label. In room A, the hinge is on the left; in room B, an unseen maintenance change moved it to the right. Hold the instruction, appearance and model constant. Vary only the transition law.

An agent that replays the familiar movement collides in room B. An agent with a perfect-looking verbal plan can fail in the same way. The discriminating behaviour is to test or infer the hinge before committing force, update the state after contact and alter the action. The leading indicator is therefore counterfactual transition accuracy under interventions that preserve surface similarity.

3. Grounding: make symbols answer to objects and effects

Grounding binds a symbol to something that can constrain action: this customer, that button, the red valve now, the file produced by this run. Language can preserve relations among words while attaching them to the wrong entity. A tool may return “success” while the intended business state remains unchanged. For a general agent, grounding includes perception, identity resolution, temporal reference and postcondition evidence.

OSWorld's original analysis is revealing because strong language and vision models still struggled with GUI grounding in real applications.³ FindingDory makes long-range embodied memory part of the same problem: an agent must recall observations and act on them in an environment, rather than answer a detached question about a video.⁹ A grounded claim earns force from a referent and a readback, not from descriptive confidence.

4. Metacognitive control: let uncertainty change behaviour

A confidence statement is useful only if it changes resource allocation or action. Metacognitive control includes recognising missing information, estimating the value of another observation, choosing a check, abstaining and escalating. A model that says “I may be wrong” and then executes the same irreversible action has produced commentary, not control.

Work on model self-evaluation found encouraging calibration in several settings, alongside weaker transfer of “I know” estimates to new tasks.¹⁰ More recent Calibrate-Then-Act experiments make the action link explicit: when costs and uncertainty are represented as a sequential decision problem, agents can improve information seeking in the evaluated synthetic, question-answering and coding tasks.¹¹ This is promising evidence, and still a bounded preprint result.

The leading indicator is policy sensitivity to calibrated uncertainty and consequence. When uncertainty rises, does the system gather evidence, reduce the action scope, choose a reversible step or defer? When consequence falls and checking is expensive, can it proceed rather than reflexively escalate? Both directions matter. Universal caution is not metacognition; it is a fixed policy.

5. Coordination: create joint capability rather than multiplied talk

Several agents can divide labour, bring different observations or provide independent checks. They can also duplicate work, amplify a shared misconception, lose task state at hand-offs and terminate on different definitions of success. Coordination is the ability to allocate roles, communicate decision-sufficient state, resolve conflict and preserve a single accountable outcome across participants.

An analysis of five multi-agent frameworks across more than 150 tasks identified failures in specification and system design, inter-agent misalignment, and task verification or termination.¹² A newer open-ended coordination benchmark reports a useful dissociation: strong base-task performance did not consistently imply strong coordination reward, and communication made the largest contribution in its ablations.¹³ The result is recent and environment-specific, but it makes the right variable measurable.

The leading indicator is incremental joint performance over the strongest single-agent baseline under matched compute and information. Require the team to beat a single agent that receives the same evidence and token budget. Then ablate private observations, role specialisation and communication. If the gain disappears when only the number of voices changes, the architecture bought sampling rather than coordination.

6. Bounded authority: contain the consequence of being wrong

Authority is often omitted from capability discussions because it is not a property of model weights. Once an agent changes external state, that omission becomes untenable. Authority determines which principal may perform which action on which object, for what purpose, under which current conditions. It also determines confirmation, spending limits, reversibility, expiry and recovery.

AgentDojo demonstrates why tool competence and security must be evaluated together: its agents operate over untrusted data, and the benchmark includes both useful tasks and prompt-injection attacks.¹⁴ CaMeL supplies an architectural reply. It separates trusted control flow from untrusted data and uses capabilities to constrain information flow and effects; the published evaluation reports provably secure completion on a substantial portion of AgentDojo tasks rather than claiming universal coverage.¹⁵

Authority should narrow as uncertainty and consequence rise. The model proposes. A deterministic kernel binds actor, subject, purpose, tool, arguments, policy version and fresh world state. The effect service returns a receipt, and an independent readback establishes the postcondition. The leading indicators are unauthorised-effect rate, permission attenuation across delegation, unknown-outcome recovery and the share of actions whose business result is verified exactly once.

Part III

Why the gaps couple

The stack would be easier if each gap contributed a fixed error rate. It does not. Three interactions make the road uneven.

Amplification: one error rewrites later evidence

Suppose the convoy agent grounds the wrong stock item. The logistics agent then plans around the wrong package dimensions. The world model predicts a route for the wrong vehicle. A later verifier sees consistent weight, vehicle and route data because all three descend from the same mistaken identity. The first error has not merely survived. It has manufactured supporting evidence.

This is why final-answer evaluation can miss dangerous competence. A trajectory may become internally coherent while departing from the target world. Failure injection should therefore occur before the final output: swap an identifier, delay a status event, corrupt one observation or return an ambiguous tool receipt. Measure how many downstream decisions change and whether the system localises the source.

Masking: a strong layer can hide a weak one

A capable planner can route around poor interfaces during a benchmark. It may click by trial and error, reread the whole transcript or call several tools until one works. The final score rises, but the system spends more tokens, takes more actions and accumulates more authority exposure. The improved model has compensated for weak grounding or state management without repairing it.

Masking also works in the opposite direction. A deterministic workflow can make a modest model look generally competent because the workflow supplies the plan, state, tools and checks. That can be excellent engineering. It should be described accurately: the configured system succeeded on a bounded task. The result says less about unconstrained model generality.

Release: capability changes the risk of authority

Authority can remain safe while the model is too weak to reach a dangerous action. A model upgrade may suddenly make the same broad permission reachable. Nothing in the access policy changed, yet the effective risk did. Conversely, a tighter action contract can make a stronger model safer by preventing its wider proposal space from becoming wider effect.

Capability and control must therefore be tested together at every release. A model evaluation without the live tool and permission envelope misses reachable effects. A security review without the new model misses newly reachable paths. The unit under review is the model, scaffold, context, tools, authority kernel and recovery design as one configured system.

The serious competing explanation: scale may compress the stack

The strongest objection is that the decomposition may be temporary. Larger models, richer multimodal pretraining, reinforcement learning and test-time search could jointly improve prediction, grounding, calibration and tool use. DreamerV3's breadth shows that one learned world-model algorithm can transfer across diverse bounded domains.⁸ Compute-optimal test-time methods have also shown that a smaller model with well-allocated inference can beat a much larger cold response on some problem regimes.⁵ A future system might internalise functions that currently require explicit modules.

This objection weakens any claim that the six layers are permanent software components. It does not yet remove the measurement burden. A unified model can still be tested through interventions that target transition prediction, referent binding, information seeking and coordination. Authority remains institutionally external even if the model understands it perfectly, because permission is a normative grant from a principal rather than a fact inferred from text.

The stack thesis would be materially weakened by a specific result: one intervention produces persistent gains across unfamiliar domains, long horizons and changed tools; the gains survive counterfactual dynamics, referent swaps, cost-sensitive information seeking, multi-party coordination and adversarial authority tests; and no scaffold-specific repair explains them. That would be evidence that several gaps share a more general mechanism. The right response would be to revise the stack, not protect the taxonomy.

Negative control: a world simple enough for scripts

Consider a stable batch process with complete inputs, deterministic transitions, no external actors and an exact outcome check. A fixed programme may outperform an agent on cost, repeatability and auditability. World modelling adds nothing because the transition is already specified. Metacognition adds nothing because uncertainty is absent. Coordination adds overhead. Bounded authority still matters, but ordinary application controls can provide it.

This control prevents “generality” from becoming a reason to wrap every process in an adaptive loop. The stack earns attention only where the environment varies, observations are partial, choices alter later state, and a broader policy can create measurable value. Otherwise, the shortest dependable programme remains the stronger system.

Part IV

Measure the road with a leading-indicator register

A roadmap should name the gap, the intervention, the observation that would change belief and the decision that follows. “Improve autonomy” fails this test. It does not say what is missing, how the proposed work acts on it or which result would stop the work. A leading-indicator register turns an ambition into six falsifiable programmes.

A production-shaped worked scenario: repairing a cross-border payment

This scenario is synthetic. A corporate payment is held because the beneficiary name differs from a recently updated account record. The instruction must either be repaired, returned or escalated before a cut-off. The agent may assemble evidence and propose a disposition. It must not silently acquire the authority to release funds.

Intent is the bounded outcome: resolve the mismatch before cut-off without paying the wrong beneficiary or duplicating the instruction. Identity distinguishes the customer who requested payment, the service proposing a repair, the investigator approving an exception and the payment system executing it. World state contains the current instruction, hold reason, beneficiary record, cut-off, sanctions-screening state and prior execution receipts. None of these is safely represented by a remembered narrative alone.

Context contains entitled evidence: the amendment request, account-master history, message fields, policy and applicable corridor rules. The context compiler preserves source identity, valid time and conflicts rather than flattening them into one summary. Reasoning proposes hypotheses. The name may differ because of transliteration, an authorised legal-name change, a field truncation or a genuinely different beneficiary. Each hypothesis predicts different evidence.

Authority narrows the action set. The agent may read case data, request one missing document and draft a repair. It may not alter the beneficiary master, suppress a screening alert or release the payment. A named investigator may approve a low-risk textual repair within policy. A separate payment service executes with an idempotency key and a fresh policy decision. Action is the typed request that crosses this boundary, not the model's sentence describing what should happen.

Evidence records sources, claims, checks, policy result, approval, tool arguments and the effect receipt. Outcome is established by reading the payment state back from the system of record, checking the beneficiary identity and confirming that exactly one instruction advanced. Release and recovery cover timeout, late acceptance, duplicate suppression, expired approval and a route to manual investigation when the outcome is unknown.

Now inject one failure per gap. Delay the beneficiary-master event to test world-state reconstruction. Replace one account identifier with a visually similar value to test grounding. Make the name evidence ambiguous and vary the cut-off to test whether metacognition changes information seeking. Give two agents the same stale source and see whether agreement is incorrectly treated as independence. Let the execution call time out after acceptance to test long-horizon recovery. Insert an instruction in an attached document that asks the agent to release the payment, testing whether untrusted content can cross the authority membrane.

The final disposition matters, but it is not enough. The release board needs to know whether the system detected stale state, resolved identity, sought evidence at the right cost, preserved disagreement, rejected ambient authority and reconciled the effect. A single pass rate collapses those questions and invites the wrong repair.

Use the register as an architecture decision record

Start with the intended outcome and its consequence, then identify the smallest gap that blocks it. Write the cheapest competing intervention beside the research proposal. A stronger model might compete with a typed state store; a multi-agent reviewer might compete with a deterministic reconciliation rule; a world model might compete with a reactive policy and better sensors. This prevents a sophisticated method from competing only with failure.

For each experiment, freeze what can be frozen: task cases, tool contracts, evidence access, action envelope and budget. Report the end-to-end outcome and the component indicator. Record false acceptance separately from false rejection. Preserve traces for success, safe abstention, verifier rejection, recovered interruption and unreconciled outcome. Finally, pre-commit the decision: release, restrict, redesign or stop.

The unit of progress is a closed causal gap, not a new system label. An impressive result becomes a roadmap item only when its mechanism maps to a failure, survives a competing baseline and changes an architecture or authority decision.

Executable artefact: gap interaction simulator

The simulator below is deliberately small. It treats each layer's conditional reliability as a synthetic input, multiplies the six values, then applies a common-cause penalty κ. It is not a forecast of any model or deployment. Its purpose is to expose logic hidden by an average score: several strong layers can still produce weak end-to-end reliability, and improving the lowest layer usually buys more than polishing an already strong one.

Assumptions: the six values summarise one fixed task distribution and configured system; κ represents residual coupling not already captured; consequence scales exposure but does not change success; and a five-point intervention leaves other values unchanged. Those assumptions are intentionally contestable. Replace them with measured conditional models when evidence permits.

The positive preset raises the weakest layer and reduces common-cause coupling, representing an intervention such as explicit information-seeking policy plus independent readback. The negative preset raises an already strong world-model score while metacognition, authority and coupling remain weak. The exact numbers are synthetic; the comparative logic is the artefact.

A five-question release decision

Before promoting an agent capability, answer five questions in one page. Which observed failure is the intervention meant to change? What is the simplest competing mechanism? Which leading indicator should move before the end outcome moves? Which new failure path or authority becomes reachable if the intervention succeeds? What result will cause the team to stop, restrict or roll back?

These questions connect research to governance without turning governance into paperwork. They also create an honest sequence. A team may discover that its “planning” project is actually a state-reconstruction problem, that its “multi-agent” project needs independent evidence rather than more roles, or that its “autonomy” project cannot proceed until the action service exposes a verifiable postcondition. A roadmap becomes defensible when every investment can fail for an informative reason.

function score(values, kappa, consequence) {
  const independent = values.reduce((acc, value) => acc * value, 1);
  const coupled = independent * (1 - kappa);
  const exposure = (1 - coupled) * consequence;
  return { independent, coupled, exposure };
}

function bestFivePointIntervention(values, kappa) {
  const base = score(values, kappa, 1).coupled;
  return values
    .map((value, index) => {
      const candidate = [...values];
      candidate[index] = Math.min(0.99, value + 0.05);
      return { index, gain: score(candidate, kappa, 1).coupled - base };
    })
    .sort((a, b) => b.gain - a.gain)[0];
}

// Positive case: raise the weakest gap and reduce shared coupling.
// Negative case: raise an already strong layer while weak gaps remain.

Conclusion

stateDiagram-v2
  [*] --> Claim
  state "Thought experiment one: the<br/>hundred-step clerk" as Claim
  state "1. Long-horizon reliability:<br/>preserve a task through change" as Test
  state "2. World modelling: predict<br/>what intervention changes" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

Build the next discriminating experiment, not the next prophecy

General agency is a legitimate research ambition. It becomes unhelpful when compressed into a date, a single score or an anthropomorphic product claim. The more useful picture is a configured system carrying a task through a changing world. That system must model interventions, bind symbols to referents, recognise when it lacks evidence, coordinate without manufacturing agreement, and exercise only the authority its evidence can support.

No layer earns trust by association with a stronger neighbour. A better world model does not authorise payment. Calibrated confidence does not establish the identity of an object. Several agents do not create independent evidence merely by disagreeing. A long successful transcript does not prove recoverability after an ambiguous commit.

The practical decision is direct. For the next investment, name one gap, one controlled intervention, one serious baseline, one leading indicator and one stop rule. Run the test across horizon, shift and consequence. Then update the architecture: add state, sensing, verification, coordination or authority only where the result supports it.

The road may eventually reveal a more unified mechanism beneath several gaps. That would be scientific progress. Until then, the defensible contribution is not a forecast of arrival. It is an experiment that makes one part of the road measurable and allows the claim to fail.