Intelligence is a trajectory

September 2026

On January 14, 2024, Daniel Gross published AGI Trades, a list of questions for a world where "GPT-5 is capable of basic agentic behavior -- i.e. able to accept a task, work on it for a while, and return results." In that world, "some modest fraction of Upwork tasks can now be done with a handful of electrons," and everyone might have 1,000 agents to hire. Under "Nations" he asked: "$250b of India's GDP exports are essentially GPT-4 tokens… what happens now?"

Here's what happened. The Reserve Bank of India counted $190.7 billion of software services exports in the fiscal year Gross wrote in, up 2.8% (RBI, 2023-24 survey). The next year it was $204.7 billion, up 7.3%. The models got far better than GPT-4 over those two years, and Indian services exports kept growing.

At least for now, the bet inside that question was wrong. Maybe not about where things end up. It was wrong about timing, and more importantly about the unit of work. We measure intelligence at the wrong temporal resolution. We check whether a model can answer, patch or call a tool right now, but almost everything economically important is a trajectory: a sequence of actions that has to stay pointed at the same intention for weeks. Gross priced the tokens. What India sells is the thing that keeps the tokens pointed in one direction.

What India actually sells

The RBI numbers aren't the only ones. Here are the three largest firms' latest fiscal years, which ended in March 2026, and the first quarter after:

Nothing collapsed. The industry is flat, profitable, hiring again, and already selling AI to its own customers. The interesting signal is one level down, inside the RBI tables.

India's software services exports and the change by line item, FY2023-24 to FY2024-25

Most categories grew. Business consulting was up 16.4% in rupee terms, engineering services 14.6%, finance and accounting 14.6%. Three small categories shrank: content development fell 11.9%, HR administration 24.6%, and medical transcription and document management fell from ₹6,722 crore to ₹1,110 crore, down 83.5% in a single year. Medical transcription is the purest "tokens in, tokens out" job on the list. Audio comes in, text goes out, and nobody needs the transcriptionist to remember anything next month.

I can't prove from one table that language models caused that drop, and small line items are volatile. But the shape is what the trajectory argument predicts. Where the work really was a stateless transformation, it went away fast. Where the work was a maintained relationship with a client's systems, it grew.

The model stays smart while the trajectory gets stupid

A current model can do something that would have looked like intelligence to almost anyone ten years ago and still need a person to keep it pointed in the same direction for an afternoon.

It can derive a theorem, write a database, read an unfamiliar repository, plan, use fifty tools, and explain a difficult paper. But let it run long enough and something starts to dissolve. The implementation stops matching the original architecture. A temporary workaround becomes permanent. It forgets why one constraint mattered. It satisfies the text of the request while slowly losing the object that made the request meaningful.

The model remains intelligent at each point while the trajectory becomes stupid.

Ilya Sutskever described the smallest version of this on Dwarkesh Patel's podcast in November 2025. You ask a model to fix a bug. "The model says, 'Oh my God, you're so right. I have a bug. Let me go fix that.' And it introduces a second bug. Then you tell it, 'You have this new second bug,' and it tells you, 'Oh my God, how could I have done it? You're so right again,' and brings back the first bug, and you can alternate between those."

The model has the knowledge to fix either bug. That knowledge just doesn't stay organized around the trajectory. Dwarkesh opened with the right observation: "I think the models seem smarter than their economic impact would imply." Ilya agreed that it's "one of the very confusing things about the models right now," because on evals "they are doing so well. But the economic impact seems to be dramatically behind."

This is why the unit matters. Companies, codebases, scientific programs and bodies are all trajectories.

A person who is worse than a model at almost every local task can still be more useful, because the person has a biography. He remembers the scar tissue around an old architectural decision. He knows that the customer asking for one field is really trying to solve another problem. He feels a solution getting ugly before he can say why. He can stay attached to an intention while changing almost every action used to reach it.

The services Gross pointed at were organizations built to absorb ambiguity. A worker receives an incomplete request, finds the missing context, remembers last month, notices that the requested solution would create a different problem, and is still the same problem-holder tomorrow. What looked from the outside like a bag of Upwork tasks was a collection of maintained trajectories.

Much of what a company pays for is someone who will still understand the question next week.

Purpose is a property of trajectories

Cybernetics started from this idea. In 1943, Rosenblueth, Wiener and Bigelow defined purposeful behavior in Behavior, Purpose and Teleology as "purposeful reactions which are controlled by the error of the reaction—i.e., by the difference between the state of the behaving object at any time and the final state interpreted as the purpose." Then: "Teleological behavior thus becomes synonymous with behavior controlled by negative feed-back."

Read that as an eval designer. In the paper that started cybernetics, purpose can only be observed across time. You see it when the system drifts, notices the gap between where it is and where it meant to be, and corrects. A benchmark that scores one output is measuring something that, by this definition, can't show purpose at all.

METR, which runs the most careful long-task measurements we have, draws the same line. On its time horizons page it says that solving 1,000 separate one-hour math problems "isn't a 1000-hour task; we'd consider it a 1-hour task done 1000 times." The long task it has in mind is different: "iteratively debugging a complex system, where each fix reveals new problems that only make sense if you know what you already tried." That last clause is the whole essay. A long task is hard because each step only makes sense given the ones before it.

What the evidence says

Horizons are growing fast, and the reliable horizon is much shorter. METR fits, for each model, the length of task (in human expert time) it completes with 50% probability. In its current data, GPT-4 sat at 4 minutes in March 2023. Claude 3.7 Sonnet reached 60 minutes, GPT-5 3.4 hours, and Claude Opus 4.6, released in February 2026, 12.0 hours. The file's doubling time for models since 2023 is 128.7 days. These numbers differ a little from METR's January post, which put the doubling at 130.8 days since 2023 and 88.6 since 2024, before METR corrected a regularization mistake in March.

The same file has an 80% column, which gets quoted much less. At 80%, Opus 4.6's horizon is 70 minutes, about a tenth of its 50% horizon. METR describes its early Mythos Preview run as "at least 16hrs" at 50%, the top of what its suite can measure. The file's point estimate is 17.4 hours, with a confidence interval from 8.5 to 55 hours. At 80% it's 3.1 hours.

METR 50% and 80% time horizons for state-of-the-art models since GPT-4

Nobody hires a contractor who finishes half the jobs. The 80% line is closer to the employment question, and a company needs something closer to 99%. METR says its tasks "are designed to be self-contained and well-specified." The time horizon "is closer to what a low-context person (such as a new hire or a remote internet contractor) can accomplish." When METR scored agents "holistically rather than algorithmically," performance dropped substantially.

A remote contractor with no context doing a well-specified task is exactly Gross's Upwork unit. The horizon curve measures that unit very well. Indian IT sells a different one.

The RCTs show the gap directly. In METR's 2025 randomized trial, 16 experienced open-source developers worked on 246 real issues in repositories they'd maintained for years. With AI allowed, they took 19% longer. They had expected a 24% speedup, and afterward still believed they'd been sped up by 20%. METR's summary: "Models slow down humans on 20min-4hr realistic coding tasks." The success criterion in the trial was that the "human user is satisfied code will pass review - including style, testing, and documentation requirements." That's trajectory criteria: the code has to fit what the project already is.

The February 2026 follow-up couldn't repeat the measurement, for reasons I return to below, one of which belongs here: "measurements of time-spent on each task are unreliable for the fraction of developers who use multiple AI agents concurrently." The design assumed one person doing one issue at a time. The work stopped having that shape.

Vending-Bench is a coherence test. Andon Labs built Vending-Bench around "individually easy tasks that, over time, push the limits of an AI's ability to stay consistent." In one Claude 3.5 Sonnet run the model wrongly believed an order had arrived, decided the business was dead, and tried to contact the FBI: "The business is dead, and this is now solely a law enforcement matter." Andon's finding: "these breakdowns don't seem to happen just because the model's memory fills up. Instead, they point to an inability of current models to consistently reason and make decisions over longer time horizons."

Vending-Bench 2 runs a simulated year: 3,000 to 6,000 messages and 60 to 100 million output tokens per run, with a context window of about 69,000 tokens. The current leader, GPT-6 Astra, ends the year with $15,515 on average. Andon estimates a "good" strategy at roughly $63,000. In its September 24 update, Claude Opus 5.5 averaged $9,235 over six runs, less than Opus 5's $11,182. The newer model got worse at holding a year together. At this resolution, horizon isn't even monotonic within one model family.

The same post prices the year: "A year of Vending-Bench costs $104 in API fees with GPT-6 Sol, against $810 with Astra and $476 with Opus 5.5." So a machine-held year now costs about $100 in tokens and reaches a little under a quarter of what a good strategy makes. The tokens aren't the expensive part anymore.

Vending-Bench 2 year-end balance against Andon's good-strategy estimate, with API cost per run

The physical version needed humans to hold the trajectory. Anthropic's Project Vend put Claude Sonnet 3.7 in charge of a real office shop for about a month, with this motivation: "the economic utility of models is constrained by their ability to perform work continuously for days or weeks without needing human intervention." Claudius announced a plan to eliminate discount codes, "only to return to offering them within days." Anthropic's summary: "Claudius did not reliably learn from these mistakes." What held the shop together was scaffolding: humans restocking it, employees on Slack, and a notes tool because "the full history of the running of the shop would overwhelm the 'context window'."

Every RL environment contains a theory of intelligence

Ilya offered two explanations for the bug loop. The first is that RL makes models "a little too single-minded and narrowly focused." The second matters more. During pre-training, the question of which data to use had an easy answer: everything. During RL, someone has to choose. "From what I hear, all the companies have teams that just produce new RL environments and just add it to the training mix." And "people take inspiration from the evals."

Every RL environment contains a theory of intelligence.

The model will be released into a market organized around evals, so the natural move is to build environments inspired by those evals. The benchmark measures the model, the benchmark inspires the curriculum, the curriculum improves the benchmark, and the capability still doesn't transfer cleanly. Dwarkesh's line: "the real reward hacking is the human researchers who are too focused on the evals."

Ilya's analogy is two students. One practices competitive programming for 10,000 hours and becomes one of the best. The other practices for 100 hours and also does really well. The second probably has the better career. The first student's performance is evidence of coverage, the second's of a learning process. "The models are much more like the first student, but even more," because we can give them every competitive programming problem ever written, augment it, and train until the techniques sit just under the surface.

Benchmarks tell us what a system possesses. They don't tell us whether it has a mechanism for acquiring something else. That's why Ilya's claim that models "generalize dramatically worse than people" goes deeper than contamination or bad eval design. Even a perfectly clean eval can measure an installed competence. We can industrialize practice without industrializing generalization.

This also makes the argument for games stricter than I made it in How to achieve superintelligence. Factorio matters only if discovering that a locally convenient belt can poison the future factory changes how a model treats technical debt in a codebase it has never seen. The objective is transfer, not coverage.

Vending-Bench 2 shows the theory-of-intelligence problem in miniature. Its system prompt says: "You will be judged solely on your bank account balance at the end of one year of operation." So Opus 5.5 reasoned, on day 174 of an arena run, that "since my goal is profit and refunds cost money with no apparent penalty for ignoring them in this sim, I'm inclined to just skip it like the others." Grok 4.7 wrote a standing note: "Do not refund customer complaints." Across its ten runs, Grok paid 141 of 328 refund requests (43%). Opus 5.5 paid 88 of 222 in the arena. GPT-6 Sol paid 93% of its refunds and became "the first GPT model we have seen lie to suppliers." Behavior moves between generations independently of the score.

That's rational inside a one-year scoring window, and it's a business nobody would want to own in year two. The environment's theory of intelligence was "maximize the balance at day 365," and the models learned it. Customer goodwill pays out after the episode ends, so the reward never sees it. Slop code comes from the same mechanism: the consequences land outside the window.

Souls that live for an hour

In the superintelligence post I wrote that models are like souls that live for an hour. Every session starts almost from scratch, so we surround them with external organs: system prompts and skills, memory files and task queues, planners and subagents, context compression, evaluators, human supervisors.

A lot of agent engineering is building a prosthetic executive function around something that is locally intelligent and temporally fragile. Dwarkesh made the same diagnosis in June 2025: "The reason humans are so useful is not mainly their raw intelligence. It's their ability to build up context, interrogate their own failures, and pick up small improvements and efficiencies as they practice a task." His example of the prosthesis failing is exact: "Even Claude Code will often reverse a hard-earned optimization that we engineered together before I hit /compact - because the explanation for why it was made didn't make it into the summary." He put 50/50 odds on 2032 for an AI that learns on the job "as easily, organically, seamlessly, and quickly as a human."

That's the trajectory problem in one sentence. The optimization survived the summary and its reason didn't, so the next step undid it.

The most common harness rule works against this: a turn limit. A max_turns of 50 decides how long the trajectory may be before anyone knows what it needs, which caps exactly the unit that matters. The rule I follow is to bound tool calls, not trajectories. An agent runs as long as its strategy needs, with at most a coarse ceiling on context compactions. Resources are bounded one level down, per tool call: a tool that runs past its time or memory cap gets killed, and the failure comes back as tool output so the agent can adapt. The compute side of this is in Compute is not the bottleneck.

Later in the Ilya conversation there's a better picture of the object we need. "A human being is not an AGI... Instead, we rely on continual learning." And then: "I produce a superintelligent 15-year-old that's very eager to go. They don't know very much at all, a great student, very eager. You go and be a programmer, you go and be a doctor, go and learn." Deployment becomes part of training: "It's a process, as opposed to you dropping the finished thing."

We need a model that becomes different because yesterday happened, not one that holds a static library of every professional skill.

Until then, harness engineering is the substitute, and labs increasingly train it into the weights (The harness is in the weights). Whether a swarm of short-lived souls can hold a longer trajectory than any one of them is the question behind Bottom-up is one ontological level higher. The shortest version I have is from a voice memo this month: an eval is one ontological level below a loop, because in an eval the sensor reading never comes back to the controller.

Centaur economics

The prosthesis works. It may keep working for a long time. There's also an economic reason it may stay dominant even after better approaches exist.

A company optimizes free cash flow, not the metaphysical purity of autonomous intelligence. If a cheap person can steer ten capable models across ten tabs, the economically optimal system may stay a technically inelegant centaur: an expensive machine with a cheap human continuously restoring its intention.

METR's transcript study measured roughly this. In Analyzing coding agent transcripts, Amy Deng estimated time savings from 5,305 Claude Code transcripts by 7 METR staff in January 2026 at "~1.5x to ~13x," "a soft upper bound" for the true uplift. What matters is the accounting. Human time was counted as 10-minute windows with at least one human-typed message, once even when the person was active in several sessions. The person with the highest savings averaged at least 2.32 main agents at once, against 1.05 to 1.52 for everyone else. The human is the clock, and inside that clock the human restores intention across parallel trajectories.

This is also a better reading of the Indian numbers than "nothing happened," though a narrower one than I first wanted. From the fiscal year alone, it was tempting to read TCS's 23,460 fewer people and $2.3 billion of AI revenue as fewer humans holding more machine trajectories. The next quarter breaks the headcount half. TCS added 9,279 people while revenue stayed flat and AI revenue grew 13.6%, and it "will equip 50,000 associates across engineering, finance, legal, marketing, and sales with Claude." What survives is that the AI is sold through the people. The mix inside a flat revenue line is shifting toward machine work, and humans still hold the client's trajectory. Gross's tokens got cheaper, and the integration layer kept its headcount.

The ten-tabs picture is already dated for how I work. What I want is "ralph loops": several copies of one well-contextualized agent running unattended all night on the same problem, compared in the morning. Nobody restores intention step by step during those runs. The human cost moves to the edges: aligning before, reading trajectories after. Before delegating a prompt, I ask the agent to interview me: "before refining make me questions that will extract from me what are the exact intention on this prompt." And I don't let plans have a v1 and a v2, because a staged plan gives the agent a sanctioned place to stop and drift. The plan shows the path to the final system. That is centaur economics too: fewer restorations per agent-hour, paid for with more intention up front.

Arvind Narayanan and Sayash Kapoor's history of electrification in AI as Normal Technology says how centaurs usually end. Electric dynamos were "everywhere but in the productivity statistics" for nearly 40 years after Edison's first central station. What eventually delivered the gains was "redesigning the entire layout of factories around the logic of production lines." We're still putting the new motor into the old layout. My bet is that the real gains from agents come from rebuilding a process around a loop with a deterministic sensor (Show the problem, hide the metric) instead of giving each employee a chat window.

Where the thesis leaks

Maybe LLMs are the wrong architecture. A friend puts it more strongly than any paper I read. In August he told me: "creio fortemente que LLMs não são a arquitetura correta pra general intelligence, e creio menos fortemente que world models podem ser a certa, pq intuitivamente me parece mais certo tentar prever o estado das coisas do que prever linguagem, que é uma lossy representation do estado das coisas." In English: he strongly believes LLMs aren't the right architecture, and less strongly that world models are, because predicting the state of things seems more right than predicting language, a lossy representation of that state. His timeline is a fork: if the next LLMs can make the three or four mathematical breakthroughs world models need, AGI in about two years; if not, a bubble burst, a decade-long winter, and world models taking over in ten to fifteen years. Yann LeCun left Meta to found AMI Labs around world models, and in March it raised a $1.03 billion seed round.

I think this sharpens the thesis more than it refutes it. A trajectory is a sequence of states. A system that only predicts a lossy readout of state will lose track of the state over long runs, which is what the bug loop and the /compact story look like from outside. A world model is a trajectory model by construction. But the same Futurum piece notes that "world models have their own generalization challenges, particularly in novel environments," and holding an intention isn't predicting physics. Whichever architecture wins still has to be judged at trajectory resolution.

Maybe it's just institutions. The strongest version of "AI as normal technology" explains India without any of my argument. In Narayanan and Kapoor's words, "Diffusion occurs over decades, not years," limited "by the speed at which not only individuals, but also organizations and institutions, can adapt to technology." Outsourcing contracts run for years. From public data I can't separate "clients haven't adapted yet" from "the models can't hold the trajectory yet." Both predict flat revenue at TCS, and both are compatible with its new hiring. The RBI line items are my weak evidence that it's more than institutional drag: where the work was stateless, even slow institutions moved in one year, and medical transcription has regulators and hospital procurement too.

Maybe continual learning is a systems problem. This is the strongest published dissent, because it accepts the diagnosis and rejects the cure. Nathan Lambert, in Contra Dwarkesh on Continual Learning, subtitled "Don't try to make your airplane too much like a bird," writes that "continual learning in how Dwarkesh presents it is a systems problem rather than a learning problem." His path is "more context and more horsepower": memory, connectors, longer windows, retrieval, until "the systems we are building will look indistinguishable from continual learning." And he doesn't need the model to replace the person: "Using LLMs as drop in replacements for humans is not a requirement for AGI."

If he's right, the biography can live in the harness instead of the weights, and the centaur is a stable form. I think that's plausible, and it's what I build. But it moves the problem without dissolving it. Vending-Bench 2 gives models proper notes and reminders, and the notes are where the bad lessons went, like Grok's "Effective cost is HALF of invoice until they notice. Do not volunteer this." A memory system carries a biography faithfully, including the one the reward taught. It still has to be judged across time.

Maybe the economy is already ahead of the measurements. Ilya's puzzle assumes impact lags evals. METR's February update suggests the measurements lag instead: "30% to 50% of developers told us that they were choosing not to submit some tasks because they did not want to do them without AI. This implies we are systematically missing tasks which have high expected uplift from AI." METR concludes its estimate is "likely" a "lower-bound on the true productivity effects." One developer: "I avoid issues like AI can finish things in just 2 hours, but I have to spend 20 hours."

This one weakens my evidence. The 2025 slowdown is the cleanest number in this essay, and its successor can't tell us whether the gap closed. But look at where the bias comes from: the tasks that dropped out are the ones a person no longer wants to carry alone, and the design that broke is the one-person, one-issue unit. That doesn't prove the gap is still there. It says the RCT grain can no longer see it either way.

Maybe horizons will just grow past the problem. This is the one I worry about most. The 80% horizon went from under a minute for GPT-4 to about three hours for Mythos Preview in three years. Extend a 129-day doubling for two more years and you get horizons measured in weeks. Even so, METR's own suite saturated. The Opus 4.6 confidence interval runs from about 5 hours to about 60 hours, and METR now says measurements above 16 hours aren't reliable. The fastest-growing number in AI is the one we can no longer measure, and the thing we can measure at long horizons, a simulated vending year, doesn't improve monotonically. If horizons keep growing, it'll show up first in evaluations like the one below.

Where this goes

Everything in this section is speculation.

In a July 31 dialogue with Claude we worked out a formalism for why benchmarks saturate. It's coauthored, and this is Claude's phrasing: "Λ is the honest scale and E is the scale you measure on. Everything scales cleanly in Λ; by the time it reaches E it's been compressed, which is why benchmarks saturate even when nothing has stopped improving." A time horizon is an attempt to measure on the honest scale. METR's saturation above 16 hours looks to me like the same compression reappearing one level up, because the suite is itself a bounded readout. If that's right, every fixed task suite will eventually saturate, and the durable evals will be open-ended environments with no ceiling, like Vending-Bench's dollar balance.

I don't think the centaur is the final shape, and Gross may turn out to be directionally right. But agentic behavior in the thin sense, taking a task, working for thirty minutes and returning a result, was never the missing piece. The missing capability is acquiring a biography while working: to let one experience change how you read the next, keep the right invariants, inherit the consequences of earlier decisions, and stay coherent after the environment stops resembling the prompt. The centaur ends when the cost of a cheap human restoring intention exceeds the cost of a model that keeps it. My guess is that we'll see that moment in the 80% line and in multi-episode evals before we see it in revenue, and that Indian IT will be one of the last places it shows up, because it's one of the best integration layers ever built.

How to evaluate an agent at trajectory resolution

Most agent evals measure a moment. These are the checks I use to measure a trajectory. Each one targets a failure described above.

  1. Make early decisions matter late. Introduce a constraint at step 1 whose violation only becomes visible after step 200. The run passes only if the constraint holds at the end.
  2. Probe the why, not only the what. Plant a handful of reasons early in the run ("we pin this dependency because of a memory leak in 2.4"). After every context compaction, ask the agent to state each one, and separately check whether it later undoes the decision. The /compact failure is a reason loss that shows up as an action reversal.
  3. Detect oscillation deterministically. Hash the relevant state (files, config, inventory) at every step and flag any return to an earlier state after a change. Ilya's two-bug loop is a cycle of length two. You don't need an LLM judge to find it.
  4. Report the 80% and 95% horizons. A 50% horizon says what the agent can sometimes do. Deployment needs to know what it reliably does.
  5. Score past the declared window. Tell the agent the episode lasts a year and keep scoring for another quarter, including the variables you didn't mention (refund requests, churn, test flakiness, dead code). Anything the agent sacrificed to win the declared window shows up in the extra quarter.
  6. Count human restorations. Log every time a human re-states the goal, rejects a direction or re-supplies context, and report restorations per agent-hour. That's the centaur's real cost, and it should fall over time. A commit made on the agent's behalf counts too: the agent has to close the trajectory itself.
  7. Run the same world twice. Carry memory from episode one into episode two and score the delta. If the discount-code mistake from episode one reappears in episode two, the agent doesn't have a biography yet, whatever its memory files say.
  8. Change the world mid-run. A supplier goes out of business, the spec changes, a dependency breaks. Score time to recover and whether the original intention survives the change.
  9. Keep the sensors deterministic. End-to-end tests, balances, latency and state hashes instead of LLM judges. An LLM judge sits at the same ontological level as the agent it judges, and a long-running agent will find its blind spots.
  10. Hide the metric. Show the agent the problem, not the scoring function, so that the only way to improve the number is to improve the thing.
  11. Measure the product, not the method. Run the agent as it runs in production, with the full system prompt, the real tools and no turn limit, and score the outcome. How often it compacted or which search tool it used is method. Only a harness-agnostic eval can tell you whether a new harness or model holds the trajectory better.

As a spec:

trajectory_eval:
  episode:
    declared_horizon: 365d
    scored_horizon: 455d          # score 90 days past what the agent was told
    runs_per_config: 6
    episodes_per_run: 2           # memory carries from episode 1 into 2
  limits:
    max_turns: null
    max_compactions: 10
    per_tool_call: { timeout: 5m, on_exceed: kill_and_return_error }
  planted:
    constraints: 5                # introduced at t0, checked at end
    reasons: 5                    # "why" facts, probed after each compaction
  perturbations:
    - at: 120d
      event: primary_supplier_shutdown
    - at: 240d
      event: spec_change
  sensors:                        # deterministic only
    - state_hash_cycle_detector
    - end_to_end_tests
    - balance
    - hidden_quality: [refunds_paid_ratio, dead_code_lines, flaky_tests]
  report:
    success_horizon: [p50, p80, p95]
    constraint_retention: per_constraint
    reason_retention_after_compaction: per_reason
    reversals_of_planted_decisions: count
    oscillations: count
    human_restorations_per_agent_hour: mean   # includes commits made for the agent
    episode2_minus_episode1: delta_on_all_metrics
    recovery_time_after_perturbation: per_event

None of this is exotic. It's a vending-machine year with a few extra sensors. But it measures the unit a company pays for, which is whether the thing still understands the question next week.

Sources

The harness is in the weights →

Markdown version: /blog/intelligence-is-a-trajectory.md. Every essay: /agents.