AI Agents for Agriculture & Farm Operationsin 2026

Almost every failed agriculture software project we have reviewed began with the same sentence: we'll integrate with whatever system the farm uses. The sentence assumes there is one. There is not. A commercial row-crop operation runs on at least four systems of record, they are owned by four different parties, and on any given day they disagree with each other about what happened in the field.

This matters more than model capability, more than the choice of framework, and more than anything on a vendor's feature list. The rest of this cluster hangs on one sentence, and agriculture is where it bites hardest: an agent can only be as autonomous as its write path allows. So before you scope a build, you have to know which of the four systems you are reading from, which one you intend to change, and who has to say yes.

The third row is the one outsiders never name, and it is often the most commercially important. The ag retailer's or co-op's ERP is where the money lives — what was sold, what was blended, what was delivered, what was invoiced, how the grain settled. Two products dominate the conversation. Agvance, from Software Solutions Integrated, carries accounting, agronomy, grain and energy modules with a Grower360 grower-facing portal. AGRIS, from Cultura Technologies, is a grain accounting and ERP platform with accounts payable and receivable, commodity management, and — this is the detail worth noticing — patronage and equity accounting as a first-class module. There is no analogue for that in generic ERP, and its presence tells you that the software was built for a cooperative, for cooperative governance, with cooperative reporting obligations. That is not a system a third-party agent gets to write to.

There is a second trap in the vendor list, and it is one that ages badly in proposals. The FMIS market looks like seven independent choices and is really a shorter list of owners. Get the ownership right before you write it into a statement of work.

Traction Ag now holds both the ex-Corteva Granular Business product, rebranded Traction Enterprise, and Conservis, acquired in December 2023. That is a market-structure fact with a direct integration consequence: a consolidation changes the roadmap, the partner program, and sometimes the API you built against. It is also a reminder that "we integrate with Granular" is not a specific enough statement to scope, because the Granular brand and Granular Insights stayed with Corteva while the FMIS product went elsewhere.

Every integration question in agriculture reduces to two: who owns the authoritative copy, and who is permitted to change it. The answers are not symmetric, and the asymmetry is the whole design constraint.

The machine platforms. John Deere runs a developer program at developer.deere.com with public documentation pages covering Operators, Assets, Files, Field Operations, webhooks, a "Precision Tech APIs" section and a changelog. Deere's own documentation describes approved applications sharing data, transferring files to JDLink-enabled machines and securely sharing files between Operations Center organizations — with approval from both the customer and John Deere — and states that becoming an ADAPT or EIC licensee is required to use the MyJohnDeere API fully. We reached that wording through a search extract rather than the page itself; developer.deere.com renders as a JavaScript shell to a fetcher, so treat it as reported rather than as a quotation. Either way the structure is unambiguous, and it is the single most useful thing to know about Deere's posture: dual consent, plus a licensing gate. That is a certified partner program. It is Rung 2 of the ladder, not Rung 1.

Deere does not publish API pricing anywhere we could reach. We are not going to estimate one. The honest statement is that pricing is not published publicly, and that a buyer must get it in writing before architecting around it.

The agronomic platforms. Climate FieldView is explicitly a partner platform, not an open API. Bayer contracted Leaf Agriculture to broker and expand FieldView's API partner connections — and Leaf is a vendor describing its own customer win, so read the framing accordingly. Bayer has also announced a FieldView partner-platform capability with RCIS on the crop insurance side, and a data-management arrangement with CLAAS TELEMATICS. The pattern to name is this: a farmer-facing platform that adds partners one bilateral deal at a time, with an intermediary selling normalisation across several such platforms. The existence of that intermediary layer is itself the evidence that the platforms are not interoperable on their own.

ADAPT is not the escape hatch people think it is. AgGateway's Agricultural Data Application Programming Toolkit is an open-source data model, a plugin manager, a set of manufacturer-supplied proprietary plugins, an open-source ISOv4Plugin that reads and writes ISOXML under ISO 11783-10, and some RESTful APIs for contextual data. The NuGet package for the ISO plugin shows an active release line — version 5.8.0, last updated July 22, 2026. ADAPT is genuinely valuable and genuinely maintained. It is also, precisely, a translation layer that depends on each OEM shipping and maintaining its own plugin. It gives you file-format parity. It does not give you a write path into a competitor's cloud.Anyone describing ADAPT as "the industry API" has not implemented against it.

And who owns the data? The industry's answer is a disclosure regime, not a rights regime. Ag Data Transparent is a not-for-profit that audits ag tech contracts against a set of Core Principles — sixteen of them as updated in 2024: Farmer Education, Simple Contracts, Ownership and Control, Collection, Notice of Changes, Transparency, Choice, Portability, Identity of the Provider, Identity of Data Partners, Disclosure and Sale Limitation, Data Retention and Deletion, Termination, Anti-Competitive Activities, Anonymization and Aggregation, and Security. Companies submit their farmer contracts along with answers to eleven questions about collection, storage, use and sharing; a third-party administrator checks that the answers match the contracts, and discrepancies must be fixed before the seal issues. The principles descend from the Privacy and Security Principles for Farm Data drafted under the American Farm Bureau Federation with commodity groups, and in February 2026 NCGA and ADT released Transparency Principles for Ag Carbon with a program contract verification process.

The integration ladder has five rungs — documented public API, certified partner program, EDI or batch file, screen scraping and RPA, and no path at all — and you are on exactly one of them per system, whatever your architecture diagram says. Agriculture sits mostly on rungs two and three, with a hard rung five at the federal boundary. Here is the honest mapping.

Two clarifications that save projects. First, wrapping a system in a Model Context Protocol server does not change your rung. MCP standardises how a tool is described and called; it does not decide whether this agent may issue this write. If the machine platform requires a licence and dual consent, an MCP server in front of it requires a licence and dual consent. MCP is a wrapper around whatever rung you were already on, and the current specification — dated 2026-07-28, with a stateless rewrite roughly eight months after the previous stable version — is a reminder that building on it means accepting that cadence.

Second, Rung 4 deserves both halves of an honest treatment. Screen scraping and robotic process automation against a grower portal or a dealer system are technically possible almost everywhere, contractually exposed almost everywhere, and structurally fragile because you are consuming a presentation layer the vendor may change without notice or versioning. The case usually cited for its legality does not say what people think it says.

In hiQ Labs v. LinkedIn, the Ninth Circuit affirmed a preliminary injunction under a sliding-scale standard. In the panel's own words on the Computer Fraud and Abuse Act leg: "HiQ has therefore raised serious questions about whether LinkedIn may invoke the CFAA to preempt hiQ's possibly meritorious tortious interference claim." A serious question is not a holding that scraping is lawful. And hiQ went on to lose the case on contract: a November 2022 summary judgment that it had breached LinkedIn's user agreement, followed by a December 2022 stipulated consent judgment of $500,000 plus a permanent injunction requiring it to cease scraping and destroy the source code, data and algorithms derived from scraped profile data. The Supreme Court's decision in Van Buren did not settle the adjacent question either — the Court wrote: "For present purposes, we need not address whether this inquiry turns only on technological (or 'code-based') limitations on access, or instead also looks to limits contained in contracts or policies."

Payback math for a back-office agent in a year-round business is straightforward: measure the current cost of a workflow, measure it again after, divide. Agriculture breaks that arithmetic in a specific way, and understanding how is the difference between a sensible first project and an expensive one.

Field operations are bounded by agronomic windows, and USDA measures those windows. NASS publishes Usual Planting and Harvesting Dates for U.S. Field Crops, in which beginning dates indicate when planting or harvesting is about 5 percent complete and ending dates when operations are about 95 percent complete, compiled by NASS field offices from historical crop progress estimates and industry specialists. And the weekly National Crop Progressreport carries "Days Suitable for Fieldwork" by state — the industry's own standard measure of how much of a calendar week was actually workable.

Those two publications let you make the economics argument without inventing a number, which matters, because the specific figures that circulate about planting-window length are per-crop and per-state and we are not going to invent an average. The denominator for ROI is not 250 business days. It is a state-and-crop-specific window measured in suitable days, which weather removes at random.

Three consequences follow, and they are the most useful thing in this article for anyone deciding what to build first.

1. 1.An integration that fails during the window cannot be fixed next sprint: The cost of a broken sync is not developer hours. It is a pass over the field that does not happen, on ground that will not be workable again for a week, in a season that happens once.
2. 2.The value of a field-critical automation is bounded by the window, and its risk is not: You capture benefit for a few weeks a year and carry failure risk for all of them. That is a poor shape for a first project, however impressive the demo.
3. 3.The administrative workflows have slack, and slack is where an agent can fail safely: Billing, reconciliation, compliance filing preparation, insurance documentation — these run on deadlines measured in weeks, not on weather. This is the strongest concrete recommendation we can make in agriculture: start there.

There is a second seasonal effect that shows up in support contracts rather than in architecture. During the window, nobody has time to learn a new interface, report a bug carefully, or work around a defect. Anything you ship must be usable by someone who is tired, in a truck, with one bar of signal. That is a design constraint, and it is the reason the deployment date matters almost as much as the design: shipping a new workflow two weeks before planting is a decision to test it under the worst conditions your users will ever face.

Connectivity compounds it. Congress directed the FCC, in consultation with USDA, to establish the Task Force for Reviewing the Connectivity and Technology Needs of Precision Agriculture in the Agricultural Improvement Act of 2018; the FCC announced its formation in June 2019 and later renewed it. Its four working groups cover mapping and analysing connectivity on agricultural lands, accelerating deployment on unserved agricultural lands, current and future connectivity demand for precision agriculture, and adoption and jobs. You do not create a statutory federal task force whose first working group is "map whether the fields have coverage" for a solved problem. Design for store-and-forward, and never let an agent assume its view of the field is current.

If you want a single example of why write paths decide what an agent can do, this is it — and it is better than any example in the other nine industries in this cluster, because the integration already exists, has existed for years, and still cannot be automated end to end.

Producers file the Report of Acreage, form FSA-578, or report through their crop insurance agent, and July 15 is the deadline for most crops. Timely filing is a condition of eligibility for crop insurance, safety-net and disaster programs, which is why it is one of the few dates on a farm calendar that nobody argues about.

Here is the part that matters. Acreage reports for approved crops are already shared electronically between FSA and RMA, eliminating duplicate entry of the common acreage information. And the producer must still contact both FSA and their crop insurance agent to complete program-specific information, validate the common information, complete maps, and sign the acreage reports.

The path is modernising, and the shape of the modernisation is instructive. On June 29, 2026, FSA announced an acreage reporting modernization pilotunder a "One Farmer, One File" effort, moving from paper maps toward an electronic interface. The pilot covers eleven counties across ten states — Tift County, Georgia; Union County, Kentucky; all counties in Maryland; Van Buren County, Michigan; Lac Qui Parle County, Minnesota; Harrison County, Missouri; Seward County, Nebraska; all counties in North Dakota; Canadian County, Oklahoma; Lancaster County, Pennsylvania; and Fisher County, Texas — and applies to spring-planted crops only in those counties, with everyone else on the standard process. The reportable data elements are crop type, intended use, acreage, field boundaries, planting dates, producer shares, irrigation practices, and prevented or failed acres. FSA Administrator Bill Beam is named in the release.

The detail an integration architect should notice: the pilot uses the farmers.gov portal for farm records and mapping, and supports precision-agriculture shapefiles from authorized third-party providers and insurance companies. That is the concrete, government-sanctioned machine path from an FMIS into a federal filing — and it is a shapefile, not an API. Rung 3, in a cluster where everyone assumes Rung 1. The release does not specify the technical architecture connecting FSA, RMA and the insurance providers, and we are not going to describe one that has not been published.

On the insurance side, crop insurance is delivered by private Approved Insurance Providers under federal reinsurance, and the data contract between RMA and the AIPs runs through the Appendix III / M-13 Handbook, which serves as the baseline for AIP Insurance Transaction Module development, with the record descriptors accessible to AIP ITM team members. RMA also publishes the Actuarial Data Master containing commodity, rate, coverage-level, plan, yield and price data, alongside an Actuarial Information Browser map viewer used by farmers, agents, AIPs and RMA. The posture to internalise: fixed-format record layouts, published by a federal agency, on an annual reinsurance-year cycle, with the descriptor detail restricted to AIP development teams. That is a batch and record-layout world, not a REST world, and the annual cycle is a hard version pin with a calendar date on it. We did not open the M-13 record layouts and will not describe specific record types or field names.

This is the most volatile fact in agriculture right now, it is live today, and it is the best available proof of a general rule about agents and regulatory values. Read it carefully, because the details are easy to get subtly wrong.

In October 2025, DOL issued a final rule amending the Adverse Effect Wage Rate methodology for non-range occupations. It split rates into two job "skill levels" defined partly by required experience, and sourced wages from the Bureau of Labor Statistics Occupational Employment and Wage Statistics by state and territory instead of USDA's Farm Labor Survey. On May 14, 2026, the U.S. District Court for the Eastern District of California denied a preliminary injunction against that methodology rule in United Farm Workers v. U.S. Department of Labor. The rule stands and the case is still pending. That is not a clean win — a denied preliminary injunction in a live case is not a merits ruling, and anyone describing it as the methodology surviving its court challenge is overstating it.

Now the part that is genuinely unusual. The 2026–27 rates took effect on two different dates depending on the state. DOL's published wage table carries the split.

The mechanism matters and is easy to garble. Kansas et al. v. U.S. Department of Labor enjoins DOL's 2024 Farmworker Protection Rule — not the October 2025 AEWR methodology rule. One of the things the 2024 rule changed was when annual AEWR updates take effect. In the seventeen enjoined states the prior 14-day lead-time regulation still governs, so the same rates land two weeks later. The wage methodology is not enjoined in those states. Only the effective-date mechanism differs. Getting that distinction wrong is the difference between a correct compliance note and a false one.

On the numbers: DOL publishes a figure of $15.96 per hour labelled "Average AEWR for Surety Bonds,"effective August 3, 2026. Use DOL's own label. It is an input to a surety-bond calculation, not a general average of what an H-2A employer pays, and we have seen it glossed as a national average across field-and-livestock occupational categories — a gloss we could not verify and will not print. Trade press has reported Skill Level I at roughly $12.31 per hour, up about 3.5 percent, and Skill Level II at roughly $16.07, up about 2.1 percent, with some states seeing much larger increases. Those two figures are trade-press reporting; DOL's own site does not publish Skill Level I/II national averages, so do not attribute them to DOL. For any specific employer, pull that state's row from the DOL table and use nothing else.

One boundary we will not cross: we did not verify current ETA-9142A or ETA-790A filing mechanics, or the present state of the FLAG system, so we are not going to describe the H-2A e-filing workflow. If a vendor demonstrates an agent filing H-2A paperwork, ask them to walk you through each screen and each signature, slowly.

Every article written about agriculture technology in the second half of 2026 will mention the FTC's settlement with Deere. Most of them will get the tense wrong, and the tense is the whole story.

On July 8, 2026, the FTC and five states — Illinois, Arizona, Michigan, Minnesota and Wisconsin — filed a joint motion for entry of a stipulated order in the FTC's monopolization case against Deere & Company, filed January 15, 2025 in the Northern District of Illinois, No. 3:25-cv-50017, before Judge Iain D. Johnston. July 8 is the date the parties signed. It is not the date of an order.The filed document reads "SO ORDERED this ____ day of _______________, 2026" — blank — above "SO STIPULATED AND AGREED this 8th day of July, 2026." The FTC's own standard framing of stipulated orders is that they carry the force of law when a District Court judge approves and signs them.

The genuinely useful part for an integration article is the FTC's underlying allegation rather than the remedy. The FTC alleged that Deere restricted the full-capability dealer tool while offering "Customer Service ADVISOR" as a lesser alternative incapable of offering the same level of functionality. That is software tiering as a competitive strategy — the same structural pattern that shows up everywhere else in this cluster, where the standards-mandated or entry tier gets you read access and the commercial tier gets you the capabilities that actually change something.

And the boundary, stated plainly because it is where the hype will go: do not claim the settlement covers APIs, telematics, or agronomic data. The reported summaries contain no such provisions and we found none. Repair-tool access is not a data-portability right. If a vendor tells you the settlement opens machine data, ask them to point at the provision, and then read it.

Everything above is constraint. Here is what is left, and it is more than it sounds — because the administrative surface of a farming operation is large, document-heavy, deadline-driven, and currently absorbed by people who would rather be doing something else.

Notice what is absent. Nothing on that list changes a machine setting, writes a rate, commits a filing, or asserts a fact to a regulator or an insurer. That is not timidity; it is where the write paths genuinely end. The boundary table below is the version we put in statements of work, and we recommend you put something like it in yours regardless of who builds the system.

The design principle underneath the table is simpler than the table: prefer reversible writes, and prefer proposals to commitments. A draft is better than a posted document. A queued recommendation is better than an executed action. A variance report is better than an automatic correction. Reversibility is the cheapest blast-radius control available and it composes with everything else in the architecture — and in agriculture it also happens to match how the work is actually supervised, because a farm office already runs on someone reviewing and signing things.

One note on scale, because it shapes who this is for. USDA's Economic Research Service, working from the Agricultural Resource Management Survey administered jointly with NASS, reports for the 2023 reference year that guidance autosteer was used on 52 percent of midsize farms and 70 percent of large-scale crop-producing farms, and that yield monitors, yield maps and soil maps were used on 68 percent of large-scale crop-producing farms. Adoption rises sharply with farm size, and small family farms with gross cash farm income under $350,000 have the lowest rates in every technology category. Use the gradient rather than the headline: it is why "farmers don't adopt technology" is false while "most farms don't have the technology" is true, and it tells you exactly which operations have the machine data an agent would read.

This is the part most proposals skip, and it is the reason this article exists. What follows is a phased path with an owner, an entry criterion, an exit criterion, and an explicit answer to the question nobody asks in the sales cycle: what do we do when this phase fails?

Two things about that sequence are specific to agriculture and worth defending. First, Phase 0 comes before Phase 1 and before anyone estimates the model work.In most industries you can start building while integration questions resolve in the background. Here, the answer to "can we write to this?" changes the product, not the schedule — and a project that assumes a write path exists and discovers it does not is a project that doubles.

Second, Phase 5 is timed against the season, not against a sprint calendar. A workflow that has not survived one seasonal peak has not been tested, however good the metrics look in February. That is inconvenient for a delivery schedule and it is the correct engineering judgement. Agriculture punishes untested assumptions once a year, on a schedule you can look up in advance.

Failure modes in this industry are mostly contract events and calendar events rather than engineering failures, which is exactly why they surprise engineering teams. Each row below pairs the failure with the signal that detects it and the rollback that contains it.

The pattern across all eight is that the agent's dependency graph extends outside your organisation, into contracts you did not draft and calendars you do not control. The practical mitigation is architectural rather than procedural: keep the draft-only path working permanently, so that any loss of privileged access degrades the system to something still useful rather than breaking it; keep every external mapping as data owned by a named person; and treat every regulatory value as configuration with a fetch-at-use path and an owner.

Prompt injection is not solved, and the people closest to it say so. The useful framing is Simon Willison's lethal trifecta: access to your private data, exposure to untrusted content, and the ability to externally communicate in a way that could be used to steal data. On whether it can be fixed, he is blunt: "we still don't know how to 100% reliably prevent this from happening." And on the vendor detection claims you will hear in every sales call, he notes that they "almost always carry confident claims that they capture '95% of attacks' or similar… but in web application security 95% is very much a failing grade."

The integration version of the trifecta, for a farm, is an agent that (1) holds a credential into a system of record, (2) reads untrusted content — a supplier invoice PDF, an email from an elevator, a scanned delivery ticket, a portal page — and (3) can act without a human seeing it first. Remove any one leg and the blast radius collapses. That is the entire practical discipline, and it is why the human-in-the-loop table earlier in this article is a security control and not just a governance nicety.

For agents with tool access to production systems, the two entries in the 2026 edition of the OWASP GenAI LLM Top 10 that deserve most of your attention are LLM03:2026 Excessive Agency and LLM10:2026 Improper Output Handling, alongside LLM01:2026 Prompt Injection itself. The mitigations cited for excessive agency are unglamorous and effective: least privilege, confirmation for high-impact actions, and tool-usage logging. Note that the 2025 numbering is different — the 2026 list puts Data and Model Poisoning at LLM05 and Unbounded Consumption at LLM06 — so if a vendor's security page cites LLM06 while describing excessive agency, they are working from last year's list.

The Model Context Protocol specification's security document is worth reading for a reason most people miss: it is entirely about confused deputies, token scope, redirect validation, server-side request forgery and consent — that is, blast radius — and contains nothing that prevents injection. Its scope-minimisation section names the risks directly, including an expanded blast radius when a stolen broad token enables unrelated tool and resource access, privilege chaining, and audit noise when a single omnibus scope masks user intent per operation. Its named anti-patterns are wildcard and omnibus scopes and bundling unrelated privileges. A security document from a standards body that is 100 percent blast-radius controls and 0 percent injection prevention is the tell.

On identity, the honest position is that this is mature inside a single vendor's estate and unsettled across estates. Microsoft Entra Agent IDgives agents identity accounts distinct from workforce, customer and workload identities, with two named modes — autonomous access using rights given directly to the agent identity, and delegated access using rights given to the user, with the user controlling which rights are delegated. Microsoft's own rationale for not simply reusing a service principal is the best available articulation of the problem: application identities carry an expectation of long-term stability, known ownership and managed lifecycle, whereas an agent might exist for minutes or be created and destroyed thousands of times a day. Practically, though: Entra governs Microsoft's estate. It does not govern your FMIS, your OEM platform or your co-op's ERP. Cross-vendor delegation that survives multiple hops has no deployed standard today.

If an agent is reading a document that arrived from outside the operation, it does not hold a credential that can change anything. If it holds a credential that can change something, it is not reading anything a stranger wrote. When both are unavoidable, a human sees the action before it happens.

— The rule of thumb we hold ourselves to on every engagement

Agriculture technology has a long history of impressive demonstrations that do not survive contact with a partner agreement. These are the signals we treat as disqualifying or near-disqualifying, and the question that exposes each one.

The general rule this cluster states out loud once: in agriculture the credible numbers come from federal statistical programs and regulatory dockets — USDA ARMS, USDA NASS, the Federal Register, DOL's published wage tables — and they are freely available. The uncredible ones come from vendors quantifying the problem they sell the fix for. When both are available on the same topic, citing the vendor is a choice.

These are the bands Frenchy Digital uses to scope agent engagements in 2026. They assume the integration and write-path assessment is in scope from the start, because in agriculture that assessment changes what gets built, not just when.

Frenchy Digital cost bands for agriculture and agribusiness AI engagements, 2026.

Senior-led delivery runs $150 to $225 per hour, and ongoing retainers run $2,500 to $9,500 per month covering model and dependency upgrades, evaluation expansion, incident response and a quarterly technical review. Every engagement carries a 30-day post-launch warranty, and you receive a written scope with a fixed-price phased proposal within 5 business days of the discovery call.

The budgeting note that surprises people: the integration substrate is largely a fixed cost, paid once and reused by everything after. The first agent pays for the FMIS connection, the file-normalisation pipeline, the identity model, the review queue and the logging. The fourth agent inherits all of it. Operations that sequence their automation get materially better economics than operations running four disconnected pilots at once — and in an industry where the working windows are narrow and the seasons are singular, sequencing is not just an economic preference. It is how you avoid finding out that three things are broken during the two weeks you needed them.

This article has argued repeatedly that the absence of a source is worth more to a reader than a confident number. That standard applies to us, so here is the list of what we could not establish while writing it.

• Whether the Deere stipulated order has been entered: As of this research pass it was filed on July 8, 2026 as a joint motion for entry and had not been entered. That single fact changes the tense of every sentence about the settlement, and the ten-year term, the 60-day reporting cadence and the baseline rollout tranche all run from entry. Check the Northern District of Illinois docket for No. 3:25-cv-50017 before you act on it.
• John Deere's API commercial terms, pricing, rate limits and the exact ADAPT/EIC licensing requirement: developer.deere.com does not render readable content to a fetcher. Everything we say about Deere's API posture is reported rather than read at origin, and we deliberately paraphrase rather than quote it.
• The FTC's own press-release text on the Deere settlement: ftc.gov returned HTTP 403 to our fetcher. The settlement terms above are corroborated secondary reporting from a law-firm alert and trade coverage, not the agency's own words as read by us.
• Skill Level I and Skill Level II national average AEWR figures: The roughly $12.31 and $16.07 figures are trade-press reporting. DOL's own site publishes the surety-bond average of $15.96 but not Skill Level I/II national averages, so we do not attribute those two numbers to DOL, and neither should anyone else.
• The technical architecture connecting FSA, RMA and the insurance providers: The June 2026 FSA release does not specify one, and we are not going to describe an architecture that has not been published. We also did not open the RMA M-13 record layouts, so we describe no record types or field names.
• H-2A e-filing mechanics: We did not verify the current ETA-9142A or ETA-790A filing mechanics or the present state of the FLAG system, so we do not describe that workflow at all.
• The current count of Ag Data Transparent certified companies: We could not obtain a current number and will not guess. If you want one, count the list yourself.
• Any independent measurement of agent integration success or failure rates: None exists that we could find, in agriculture or anywhere else. That is why this article contains no adoption or success percentage for agent projects, and why every business case has to rest on a baseline you measured yourself.
• Any average planting-window length: It varies by crop, state and year. NASS publishes the underlying per-crop, per-state handbook; we did not extract values and will not print an average that does not exist.
• Prompt injection remains unsolved: Any agent reading supplier documents, portal pages or inbound email is exposed. We reduce blast radius through capability scoping, deny-by-default arguments, reversible writes and human approval gates. We do not claim to eliminate the risk, and nobody who does should be believed.
• We are not agronomists, insurance agents or immigration counsel: Frenchy Digital builds software. Agronomic decisions, insurance attestations, crop-program eligibility and H-2A compliance belong to qualified professionals, and our systems feed their judgement rather than substituting for it.

None of this argues against building. It argues for building the measurement alongside the system, choosing one administrative workflow with real slack in it and a named owner, and being honest inside your own organisation about which claims in this category have evidence behind them and which do not. In agriculture that honesty is worth more than usual — because the season does not offer a second attempt, and a business case resting on a number nobody can source is a business case that fails in the one week you needed it to hold.