Progressively Autonomous Still Waits: The Approval Round Three Operators Do Not Price

Sinopec's interim results of 23 August 2026 report that the company "further carried forward" its AI Plus initiative "with the launch of the industry's first digital expert, namely the 'Fenghuo' industrial AI agent", and that the capabilities of its Great Wall large model "further improved" [1]. On 1 September YPF published the renewal of its agreement with Corva, which the release describes as one of the central technology platforms of its RTIC de Perforación y Workover, the real time centre for drilling and workover, and set out the next stage as advancing towards "operaciones progresivamente autónomas mediante advisors, analítica predictiva, alertas inteligentes, automatización, copilotos digitales y agentes capaces de asistir procesos operativos de creciente complejidad": progressively autonomous operations, by way of advisors, predictive analytics, intelligent alerts, automation, digital copilots and agents able to assist operational processes of increasing complexity [2]. Two days later PETRONAS, through Malaysia Petroleum Management, announced an agreement with Iraya Energies to add agentic AI to the myPROdata upstream data platform, where the enhanced platform "will integrate" its data with agentic capabilities that "can autonomously analyse and connect multiple data sources" [3].

Three operators, three statements of capability, and not one rate. None of the three says what share of its agent's actions still stops for a person, how long that person takes to pick one up, or how many such stops a cycle contains. That is not a complaint about the releases. Two are partnership notices and one is a results announcement, and none of them claims to be an operations study. It is the reason the sentence in the title is worth saying, because those three numbers, and not the model, are what set the calendar time of an agentic loop once the loop works at all.

Three statements, read exactly

What each one does and does not say is the whole of the evidence, so it is worth reading to the letter.

Sinopec's is a launch. The release names two things, the Fenghuo agent and the Great Wall model, and quantifies neither: there is no count of users, sites, wells or workloads attached to either anywhere in the document [1]. The nearest related claim in the same section is that breakthroughs were made in "synergistic oil flooding theories and intelligent drilling methods", which is a research sentence with no number on it either [1]. Nothing in the release describes the agent as running in production, and nothing here should either.

YPF's is the most explicit of the three about direction, and the least about arrival. The agreement "entra ahora en una nueva etapa", enters now a new stage, oriented towards progressively autonomous operations [2]. Read the list of what delivers that stage and every item on it is assistive by construction: advisors, intelligent alerts, digital copilots, and agents "capaces de asistir", able to assist. The release also describes what the centre already does, and it is a supervision arrangement: specialists and field teams can see performance in real time, anticipate risk events, reduce non-productive time and decide faster [2]. A person is in that sentence. There is no figure of any kind in the release, and no date by which any of it is expected.

PETRONAS is the one release that names the substrate, and it is the one that says least about deployment. myPROdata is described as a web-based platform giving access to subsurface and surface data on exploration blocks, discovered resource opportunities and producing fields; the agreement is that the enhanced platform "will integrate" that data with agentic AI [3]. Nothing is described as live. The RM50 billion to RM60 billion of annual upstream investment, and the shortening of the journey from discovery to first hydrocarbon from 100 months to 50 months, are quoted in the release as PETRONAS' ambition in a statement by Malaysia Petroleum Management's senior vice president, and they are not outcomes of this agreement [3].

So the common ground is real and it is narrow. Three operators are buying assistive autonomy, all three describe a system that proposes and a person who remains in the arrangement, and none of them prices the person.

What an approval costs a loop, from our own record

We have built two of these loops and recorded what the human side of them cost.

For a national oil company in Southeast Asia we built an agentic interpretation loop in which a foundation model proposes a full interpretation, a review agent scores every inline and crossline pick for geologic plausibility, and only the low-confidence segments go to a geoscientist [4]. Interpretation went from six weeks per prospect to four days, with 72% fewer geoscientist hours [4]. The numbers that matter here are the smaller ones. On the first production prospect the geoscientist spent four hours reviewing flags, made corrections on 11% of the proposed picks, and signed off; the prospect took nine days end to end. By the third, cycle time was four days and review was down to ninety minutes, and six months in the flag rate had fallen from 11% to 4% [4].

For a global supermajor we built an agentic MLOps loop around an existing model estate: monitoring, data reconciliation, retraining and validation agents in sequence, with a mandatory physics gate and a one-click geoscientist approval before any model is promoted [5]. Six-week retrain cycles became overnight jobs, about 40 times faster, and production-forecast accuracy across roughly 40 producing assets lifted 22% [5]. That approval gate was not an oversight we tolerated. It was added after one retrain passed every statistical test and violated basic reservoir physics on two wells, and the brief from the client's VP of Digital Subsurface had been explicit from the start: models that keep themselves current, with the expert as approver rather than operator [5].

Two things in that record decide the arithmetic below. The first is that the gated share is small and it falls: 11% at the start, 4% after six months [4]. The second is easy to miss and it is the one that pays. Our record has the geoscientist reviewing the flags and signing off in a single four-hour sitting, not being interrupted once per flag. The loop ran ahead, pooled what it wanted approved, and presented it once.

The arithmetic

Take a cycle of $N$ agent actions, a share $g$ of which stop for a human. Actions are whole, so the count that stops is

$$k \;=\; \lceil g N \rceil$$

FormulaLaTeX

Copy LaTeX

which is at least one for every share above zero. The loop pools up to $b$ of them into one approval round, so the number of rounds the cycle waits through is

$$R \;=\; \left\lceil \frac{k}{b} \right\rceil$$

FormulaLaTeX

Copy LaTeX

Each round costs the cycle the approver's wait $W$, the time from raising the round to the approver picking it up, and each approval costs the approver's own handling $h$. With $M$ the agent's own run for the whole cycle,

$$T \;=\; M \;+\; \left\lceil \frac{\lceil g N \rceil}{b} \right\rceil W \;+\; \lceil g N \rceil\, h$$

FormulaLaTeX

Copy LaTeX

Three properties follow with no further assumption, and they are the whole of the finding.

The cycle time is a staircase in the gated share, not a slope. Between risers, lowering $g$ by one action saves $h$ and nothing more; at a riser it saves $W + h$. Since $W$ is measured in days and $h$ in minutes, essentially all of the saving is concentrated in the risers, and there are $\lceil gN/b \rceil$ of them.

The last riser never disappears. For every $k$ from 1 to $b$, the ceiling is 1, so

$$T \;\ge\; M + W + h \qquad \text{for every } k \ge 1$$

FormulaLaTeX

Copy LaTeX

and the only way to remove that last $W$ is to gate nothing at all. Progressive autonomy walks $g$ towards zero and hits this floor long before it arrives.

The agent is in one term. $M$ appears once and additively, so as the agent gets faster the cycle time converges on $RW + kh$, which contains no property of the agent whatsoever. That is Amdahl's argument with the serial fraction relabelled: the part that cannot be automated is not a part of the computation, it is the person the computation waits for.

The bench

The exhibit below computes that surface rather than drawing it. Left to right is the share of agent actions gated on a human, 0 to 100%. Back to front is the approver's wait per round, one hour to five days, on a log axis. Height is end to end cycle time, also on a log axis, whose top is the six-week manual cycle we recorded. The amber sheet is the overnight anchor. The white curtain is the gated share you set, and the labels follow it.

The bench opens on our own record and says on the plate which parts of it are recorded and which are inferred. Recorded: six weeks by hand, overnight with approvers, 400 inlines in the manual first pass, 11% of picks corrected, four hours of review [4][5]. Derived: 11% of 400 is 44 flags, so the handling rate is 5.45 minutes an approval, and the pooling opens at 44 because our record reads as one pass over all of them. Assumed, and labelled as such: that overnight means 16 hours, and that the approver waits four hours. Our record does not split the nine-day first prospect into agent time and approver wait, so the agent's own run is not measured either; it is derived on the plate as the overnight anchor less one four-hour round less the four hours of handling, which is eight hours. Four of those are sliders: the gated share, the approver wait, the agent's run and the pooling. The overnight anchor is fixed at 16 hours.

gated share of agent actions, 0 to 100%: 11.0%approver wait per round, 1.0 h to 5.0 days: 4.0 hagent run for the cycle, 6 min to 10.0 days: 8.0 happrovals pooled into one round, 1 to 400: 44

What the bench shows

At the opening settings the cycle reads 16.0 h, which is where the derivation puts it, on one approval round. The approver floor, which is the cycle with the agent's run taken to zero, reads 8.0 h, and overnight is within reach.

Now drag only the approver wait, from four hours to two days. The cycle goes to 2.5 days, the approver floor to 2.2 days, and the overnight verdict flips to out of reach. Nothing about the agent changed. Then drag the agent run down to its minimum of six minutes: the cycle falls to 2.2 days and stops there, while the approver floor stays at 2.2 days and the verdict stays out of reach, because the expression behind those two does not contain the agent's run. An agent eighty times faster takes about eight hours off the cycle and nothing at all off the floor.

Now take the gated share down, which is what progressive autonomy means. From 11% to 10% the cycle reads 2.5 days at both, because both counts sit inside one round and four fewer approvals is 22 minutes of handling. Take it to 1%, four actions gated, and the cycle is 2.3 days on a floor of 2.0 days. Drag it as far left as it goes without reaching zero and you are down to a single gated action; neither figure moves. Still 2.3 days on a floor of 2.0 days, and the one-action readout now agrees with the hero, because that is where you are. The whole of progressive autonomy from 11% gated to one action gated is worth about four hours here. The last gate costs two days on its own.

Push the other way and the staircase becomes visible. Put the approver wait back to four hours and drag the gated share up: at 25% the rounds readout goes to 3 and the cycle to 29.1 h, at 50% to 5 rounds and 46.2 h, at 100% to 10 rounds and 3.5 days. The risers are where the time goes, and the pooling slider is the control that moves them. At 100% gated with a two-day approver, pooling everything a cycle flags into one round gives 1 round and 3.8 days; pooling one approval at a time gives 400 rounds and 114.5 weeks, at which point the surface has flattened against the top of the block, because the loop now takes longer than the six weeks of hand work it replaced. Same agent, same gated share, same approver. The difference is whether the loop interrupts the person 400 times or once.

That is the shape of the argument. Autonomy and pooling both reduce rounds, but they are not interchangeable: pooling collapses many rounds into one and can do it today, while autonomy grinds the count down towards a floor of one that it cannot pass. And under that floor sits the approver's own clock, which is a rota and an escalation policy, not a model.

What the bench is not

The two anchors are ours and the rest of the surface is the reader's, and it is worth being exact about which is which.

The 400 actions are the inlines our record describes a geoscientist scrolling through in the manual first pass [4]. Our record does not state how many picks the agent proposed, so 400 is a stand-in for the action count and the crosslines the review agent also scores would raise it. It does not matter to the finding: the floor $M + W + h$ contains no $N$, and doubling both the action count and the pooling leaves every round unchanged. What $N$ sets is the tread of the staircase, not its floor.

The handling rate of 5.45 minutes an approval is four recorded hours divided by a count that is itself 11% of 400. It is the softest number on the plate and it is also the one that matters least, because it is minutes against a wait measured in days. Set it aside entirely and every conclusion above survives.

The approver wait is the number we do not have and neither do the three releases. Ours is not recorded, so the bench opens at an assumed four hours and says so on the plate. That is precisely the number this piece is asking for, and it is the one an operator can measure this week without a model in sight.

Three questions for a progressively autonomous loop

What share of the agent's actions stops for a person today, and what was it a quarter ago? Ours went from 11% to 4% in six months on one loop [4], which is the trajectory the word progressive describes. The share is the easy number and it is the one that will be quoted.

How long does an approval wait before someone picks it up, and is that a queue or a rota? This is the number none of the three releases contains and the number the arithmetic above is most sensitive to. An operator that can answer it in hours has a loop that can run overnight; one that answers it in days does not, whatever the agent does.

How many times per cycle does the loop stop, and can it run ahead? A loop that pools its approvals into one review pass, which is what our own delivery did [4], pays one wait. A loop that raises them one at a time pays one wait per gated action, and at any realistic approver that is the difference between a cycle in days and a cycle in months. Agents "capaces de asistir procesos operativos de creciente complejidad" [2] will raise more of these, not fewer, as the complexity they assist with grows. The design question is not how autonomous the agent is. It is how many times it is allowed to interrupt the person, and how fast that person answers.

Key takeaways

1. 1Sinopec reports launching the Fenghuo industrial AI agent, YPF and Corva set the next stage of YPF's RTIC de Perforación y Workover as operaciones progresivamente autónomas delivered through advisors, copilots and agents that assist, and PETRONAS says myPROdata will integrate agentic AI. All three are capability statements; none states a gated share, an approver wait, or a count of stops per cycle.
2. 2On our own agentic interpretation loop the geoscientist corrected 11% of the proposed picks, spent four hours reviewing flags, and signed off in one pass; six months later the flag rate had fallen to 4%. On our agentic MLOps loop a one-click approval gate was added after a retrain passed every statistical test and violated reservoir physics on two wells.
3. 3Approval rounds are a rounded-up count, so cycle time is a staircase in the gated share rather than a slope: between risers a lower gated share buys back only the approver's handling, in minutes, while each riser is a whole approver wait, in days.
4. 4The last riser never disappears. While a single action is gated the cycle still carries one whole wait, so at 10% gated and a two-day approver the approvals alone cost 2.2 days and overnight is out of reach at every agent speed. Dragging the agent run from eight hours to six minutes moves that cycle from 2.5 days to 2.2 days and does not move the floor at all.
5. 5Pooling is the lever that is available now. At 100% gated with a two-day approver, one pooled round is 3.8 days and one approval at a time is 400 rounds and 114.5 weeks, longer than the six weeks of hand work the loop replaced. Same agent, same gated share, same approver.
6. 6The number to ask an operator for is not how autonomous the agent is. It is how long an approval waits before someone picks it up, and how many times a cycle stops.

References

[1] China Petroleum and Chemical Corporation (Sinopec Corp). Press Release: Sinopec FY2026 Interim Results, Sinopec Achieves Solid Operating Results in the First Half of 2026. EQS Newswire, 23 August 2026. https://www.eqs-news.com/news/corporate-news/en-press-release-sinopec-fy2026-interim-results/514ef82e-238a-4aa2-9f2d-32f1161d68b0_en

[2] YPF. YPF y Corva renuevan un acuerdo clave para el RTIC de Perforación y Workover. YPF Hoy, 1 September 2026. https://novedades.ypf.com/ypf-corva-renovacaciondelacuerdo.html The release is YPF's own; Corva is the platform partner named in it. Quotations are from the Spanish original and the translations are ours.

[3] PETRONAS. PETRONAS Advances Malaysia's Upstream Data Platform with Agentic AI to Accelerate E&P Investment. 3 September 2026. https://www.petronas.com/media/media-releases/petronas-advances-malaysias-upstream-data-platform-agentic-ai-accelerate-ep

[4] EarthScan. From weeks to days: an NOC's agentic seismic-interpretation loop. https://earthscan.io/case-studies/noc-agentic-seismic-interpretation-loop

[5] EarthScan. Agentic MLOps: From 6-week retrains to overnight. https://earthscan.io/case-studies/agentic-mlops-six-week-retrains-to-overnight