FUTURE WARFARE | The Autonomy Spectrum

Introduction: The Loop Is Breaking

In the early days of drone warfare, the chain of custody for lethal force was clear: a human operator, thousands of miles from the battlefield, watched a video feed, consulted a legal advisor, received authorisation from a senior officer, and pressed a button. That model, sometimes caricatured as ‘PlayStation warfare,’ at least maintained an unambiguous human in the trigger chain. The process was slow, cumbersome, and politically controversial — but it was accountable.

That clarity is dissolving. As of early 2026, a U.S. Air Force colonel at the AFA Warfare Symposium described watching Anduril’s YFQ-44A autonomous combat aircraft switch between two entirely different AI mission software systems — mid-flight, without landing, without a human touching the controls. Meanwhile, Maven Smart System, now deployed to more than 35 military services and combatant commands with over 20,000 active users, is fusing satellite imagery, signals intelligence, and drone feeds into targeting recommendations at a pace that compresses human decision cycles from hours to seconds. And in eastern Ukraine, three-person teams are simultaneously managing FPV strike drones, electronic warfare jammers, and UGV logistics convoys from a minivan — each system executing increasingly autonomous sub-tasks that would have required dedicated battalions a decade ago.

The question that defines this transformation is not whether AI should be in the kill chain. It already is. The question is: where, precisely, does the human still sit in that chain — and whether the physical and cognitive demands of modern combat are making meaningful human oversight structurally impossible.

I. The Three-Tier Model

Human-in-the-Loop, Human-on-the-Loop, and Full Autonomy

The defence community has developed a vocabulary for the spectrum of human control that sits between fully remote-piloted systems and fully autonomous weapons. At the most controlled end, ‘human-in-the-loop’ systems require an operator to authorise every individual engagement. At the least controlled end, fully autonomous weapons select and engage targets without any real-time human input. Between those poles sits the most contested terrain in contemporary military ethics and procurement: the ‘human-on-the-loop’ category, where a machine acts but a human retains the theoretical ability to intervene.

The most instructive current example of the transitional grey zone is Russia’s Marker unmanned ground vehicle. Operated by Android Technics, the Marker is designed to patrol and engage with a degree of independent target recognition — yet its developers maintain that lethal decisions require human confirmation. In practice, the distinction blurs rapidly in a high-jamming, communications-degraded environment: if a machine identifies a target and a human cannot override within the engagement window, the practical effect is full autonomy, regardless of the nominal doctrine.

The same ambiguity surrounds Israel’s Harop loitering munition, Turkey’s Kargu-2 — reportedly the first autonomous weapon to engage a target in combat, in Libya in 2020 — and the expanding roster of counter-drone systems worldwide that operate in fully autonomous mode once activated. The U.S. Navy’s Phalanx CIWS has been firing autonomously at incoming threats since the 1970s. Nobody calls it a ‘killer robot.’ The definitional ambiguity is not an accident: it is the space in which states operate to avoid treaty constraints.

II. Agentic AI: The New Frontier

From Rule-Following Automation to Goal-Directed AI in the Kill Chain

The term ‘agentic AI’ describes systems that act independently on stated goals — reasoning across multiple steps, using tools, and making sub-decisions — rather than simply executing pre-programmed rules. In the consumer context, this is the difference between a calculator and an AI assistant that plans your trip. In the defence context, the difference is between a proximity fuse and an AI that synthesises intelligence from 150 sources, identifies time-sensitive targets, suggests optimal munitions, and recalculates strike windows in real time.

That second scenario is not a future state. As of 2026, it describes the operational profile of Palantir’s Maven Smart System in active theatre deployments. The Pentagon originally launched Project Maven in 2017 as a computer-vision tool to scan drone surveillance footage. 

By May 2025, the contract ceiling had been raised to $1.3 billion through 2029 — up from $480 million — as combatant commands reported sharply increased demand. NATO adopted Maven Smart System for Allied Command Operations in April 2025, with the system providing ‘intelligence fusion and targeting, battlespace awareness and planning, and accelerated decision-making.’ By September 2025, the NGA director stated that by June 2026, Maven would begin transmitting “100 percent machine-generated” intelligence to combatant commanders.

Alongside Maven, the Pentagon has built GenAI.mil — a generalised AI platform accessible to every military and civilian Department of Defense employee. By December 2025, xAI’s Grok models were being integrated into the platform at a classification level allowing handling of sensitive controlled information. A poster in Pentagon hallways instructed employees the new AI tool was available and they were ‘highly encouraged’ to use it.

The implications for the kill chain are profound. When an AI system like Maven fuses satellite imagery, SIGINT, and drone feeds and surfaces targeting recommendations faster than a human analyst can read a situation report, the nominal ‘human in the loop’ becomes an approval node in a process they did not design, cannot fully audit, and may not have time to meaningfully interrogate. The distinction between human control and human endorsement becomes very thin indeed.

This tension burst into public view in February–March 2026 through the clash between Anthropic and the Pentagon over the use of AI in autonomous weapons. Anthropic, whose Claude model had been integrated into Maven Smart System via Palantir, refused to permit its AI to be deployed for fully autonomous lethal weapons without human oversight. Secretary of Defense Hegseth argued the Department could not be constrained by a vendor’s internal safety policies. The episode crystallised the central question of the autonomy debate: who ultimately governs the deployment of AI in the kill chain — the procurement agency, the platform developer, or the AI model provider?

Source: DefenseScoop, May 2025; Breaking Defense, April 2025; Project Maven Wikipedia, March 2026; TechCrunch, February 2026

III. GPS-Denied Navigation

How Shield AI’s Hivemind and Others Achieve Autonomous Flight in Electronic Warfare Environments

One of the most consequential technical realities of the current autonomy landscape is that GPS denial — the jamming or spoofing of satellite navigation signals — is the default condition in high-intensity peer conflict, not the exception. Russia deploys sophisticated GPS jamming across eastern Ukraine as standard electronic warfare practice. Any autonomous system that depends on satellite positioning to navigate is, in that environment, neither autonomous nor effective.

This has driven an urgent research and development effort in what practitioners call ‘GPS-agnostic’ or ‘comms-denied’ navigation — the ability of an autonomous system to perceive, plan, and act without any external signal input. The leading commercial implementation is Shield AI’s Hivemind platform, described by the company as an AI autonomy software stack that enables unmanned systems to conduct complex missions in GPS- and communications-denied environments.

Hivemind’s architecture achieves this through a combination of inertial measurement units, on-board sensor fusion, edge AI inference, and pre-trained environmental models that allow the system to navigate by reference to its own sensors rather than external signals. The system has been deployed across 15 aerial platforms, including the U.S. Navy’s BQM-177A target drone, Airbus’s H145 helicopter, and — in its most significant military deployment — as one of two competing mission autonomy software stacks for the U.S. Air Force’s Collaborative Combat Aircraft programme.

In August 2025, Shield AI demonstrated Hivemind’s counter-UAS application in a joint exercise with Airbus at Andøya, Norway, where a Hivemind-powered DT25 drone autonomously detected, tracked, and pursued a live-flying adversary aircraft — including during simulated GPS jamming — by fusing inertial measurement data with radio-based positioning and applying latency-compensated filters to maintain targeting when primary sensors degraded.

The flip side of GPS-denied navigation is also significant: an autonomous system that does not depend on satellite signals cannot be trivially jammed or spoofed by adversary electronic warfare. The technology that makes Hivemind resistant to GPS denial is the same technology that makes swarms composed of Hivemind-enabled drones highly resilient to the jamming countermeasures that currently represent the primary defence against small UAS.

Source: Shield AI / Airbus Defence and Space, August 2025; Shield AI World Defense Show 2026 interview; Military Embedded Systems, 2025

IV. Edge AI vs. Cloud Intelligence

On-Board Processing, Latency, and the Architecture of Lethal Decisions

The autonomy spectrum is not only a question of human oversight — it is also a question of where the computing happens. The distinction between ‘edge AI’ and ‘cloud intelligence’ is central to understanding how modern autonomous weapons systems actually work, and where the meaningful decision points reside.

Cloud intelligence models are powerful, data-rich, and updatable, but depend on communication links that can be jammed, severed, or subject to latency. In the time it takes to transmit sensor data to a cloud server and receive back a targeting recommendation, a high-speed drone may have travelled hundreds of metres. For slow, deliberate intelligence tasks — route planning, target identification from surveillance imagery, logistics optimisation — cloud AI is appropriate and effective. Maven Smart System operates in this domain.

Edge AI, by contrast, processes data on the platform itself, using compact neural network models optimised for low-power embedded processors. In the context of drone warfare, this means a small computer vision model running on a low-cost processor inside the drone that makes real-time perceptual judgements — ‘is that a tank?’, ‘is that a civilian vehicle?’ — without any communication to an external system. Ukraine’s FPV drones increasingly incorporate this capability; Russian fibre-optic drones, immune to radio-frequency jamming, rely on it entirely.

The architectural significance of this distinction was articulated clearly in the context of the Anthropic-Pentagon dispute. As one analyst observed, even if a frontier AI model is not directly involved in the final trigger decision, it may be running every decision in the chain leading up to that point — with no guarantee a human meaningfully oversees the final step. The edge-cloud architecture of modern autonomous weapons means that the ‘human-in-the-loop’ may be present for cloud-based mission planning but entirely absent from the edge-AI terminal engagement.

Source: TechCrunch analysis, February 2026; Shield AI Hivemind Architecture documentation, 2025

V. Mid-Flight AI Software Swaps

The YFQ-44A Demonstration: Mission Autonomy as an App Layer

On 24–26 February 2026, the U.S. Air Force crossed a significant technical and conceptual milestone in its Collaborative Combat Aircraft programme. Anduril’s YFQ-44A Fury — one of two fighter-class autonomous wingmen commissioned under the CCA Increment 1 contract — flew with two entirely different mission autonomy software systems in a single sortie, switching between them mid-flight without landing.

The sequence was as follows: the aircraft took off and autonomously navigated to a designated point over the Mojave desert, where Shield AI’s Hivemind mission autonomy software was activated to complete a series of test cards representative of future combat mission profiles. Following completion of those evaluations, the aircraft transitioned seamlessly to Anduril’s own Lattice for Mission Autonomy software, repeated the same test points, and returned safely to base. No hardware was modified. No redesign was required. The switch was purely a software event, enabled by the Air Force’s Autonomy Government Reference Architecture (A-GRA) — a software framework designed to standardise interfaces between aircraft systems and autonomy software, preventing proprietary lock-in.

The strategic significance extends well beyond the technical achievement. The demonstration confirmed that mission autonomy can be treated as ‘software as a service’ — an upgradeable, replaceable, competitive layer that sits atop a stable hardware platform. This is exactly the model that has made consumer technology iterate so much faster than traditional hardware procurement. An Air Force that can upgrade the cognitive capability of its autonomous wingmen via a software patch — rather than a multi-year platform redesign programme — can close the innovation loop against adversaries who operate on the same principle.

The second autonomy system — Anduril’s own Lattice for Mission Autonomy — is not currently under consideration by the Air Force for CCA Increment 1, which has selected Shield AI’s Hivemind as the designated autonomy provider. But its successful demonstration in the same sortie signals that Anduril is building competitive optionality: a hedge against programme changes and a proof of the modular architecture it has invested in across all its platforms.

The parallel from the General Atomics YFQ-42A side of the programme is equally revealing: that aircraft flew earlier in February with Collins Aerospace’s Sidekick mission autonomy software — a separate AI provider, a separate platform, the same conceptual architecture. The Air Force has designed CCA from the ground up to be a competitive software ecosystem rather than a single-vendor proprietary system. The result, if it delivers on its ambition, is an autonomous combat aircraft whose intelligence can be updated faster than the adversary can adapt to it.

Source: Anduril press release, February 26, 2026; Air & Space Forces Magazine, February 28, 2026; The Aviationist, March 3, 2026; AFA Warfare Symposium panel, February 2026

VI. Ethical and Legal Red Lines

UN Resolution 166–3, DoD Directive 3000.09, and the Regulation Gap

The international governance framework for lethal autonomous weapons systems (LAWS) has moved faster in the past 24 months than in the preceding decade — but it remains decisively behind the technology. In December 2024, the United Nations General Assembly passed a resolution endorsing a two-tier LAWS governance framework by a vote of 166 to 3, calling for regulatory monitoring of some systems and treaty bans on others. The three dissenting votes were Russia, North Korea, and Belarus.

The resolution is non-binding. It does not define LAWS with the precision required to create enforceable treaty law. And the states most actively developing autonomous weapons — the United States, China, Israel, Russia, and Turkey — have strong incentives to exploit definitional ambiguity rather than resolve it. The UN Secretary-General has called for a legally binding treaty prohibiting LAWS from operating without human oversight by 2026; there is no realistic prospect of that deadline being met.

At the national level, the most consequential policy document is U.S. Department of Defense Directive 3000.09, originally issued in 2012 and updated in 2023. Directive 3000.09 does not ban fully autonomous weapons; it requires that senior DoD leaders conduct additional review of autonomous systems before they enter the acquisition pipeline and before they are deployed. It also requires that all autonomous weapons systems be ‘designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force’ — language that is deliberately flexible and does not mandate manual human control of every shot.

What is remarkable about the current regulatory landscape is how much legitimate operational space Directive 3000.09 already creates for highly autonomous systems. Defensive systems like the Phalanx CIWS — which fires autonomously at incoming threats — are explicitly excluded from the additional review requirement on the grounds that they are purely defensive. Semi-autonomous ‘fire-and-forget’ weapons are also excluded. The vast majority of systems operating in the grey zone of partial autonomy are never subject to the elevated scrutiny the Directive was designed to provide.

The Geneva Convention’s application to AI targeting raises equally unresolved questions. The principles of distinction (between combatants and civilians), proportionality (collateral damage versus military advantage), and precaution (taking all feasible measures to avoid civilian harm) are all contextual judgements that require human moral reasoning. Whether an AI system can satisfy the legal standard for ‘precaution’ when it makes a targeting decision in milliseconds without the contextual knowledge that a human operator would bring is a question no court has definitively answered — and no treaty has yet been designed to force.

Source: DoD Directive 3000.09 (2023 update); UN General Assembly resolution A/RES/79/L.69, December 2024; Congress.gov CRS Report IF11150; Holland & Knight FY2026 NDAA Analysis, December 2025

VII. 2026 Updates: When Combat Tempo Exceeds Human Speed

The Physical Impossibility of the Human-in-the-Loop at Machine Speed

The theoretical debate about whether humans should remain in the lethal decision loop is increasingly being decided by a practical constraint: in the fastest combat scenarios of 2026, a human cannot respond in the engagement window. A drone swarm approaching a military installation at speed, a hypersonic missile in the terminal phase, or an adversary drone inside the perimeter of an air base — all present threat timescales measured in seconds, not minutes. In each case, an automated defensive response is not a preference; it is the only physically possible response.

This is the argument the Pentagon made to Anthropic in the months before the February 2026 confrontation. As Pentagon CTO Emil Michael explained, scenarios like the activation of defensive lasers to intercept incoming drones threatening a military base — where an automated response could protect personnel far faster than human operators could act — are cases where full autonomy is not a philosophical preference but an operational necessity. Anthropic countered that these specific scenarios could be addressed case-by-case while preserving a general prohibition on AI-powered lethal autonomy. The Pentagon rejected that approach as operationally unworkable.

The pressure is equally acute in offensive swarm operations. Auterion’s December 2025 demonstration of the first multi-manufacturer combat drone swarm — where a single operator directed FPV platforms and fixed-wing loitering munitions from different manufacturers as a single coordinated force — illustrated the direction of travel: ‘we’re watching the battlefield evolve from manned platforms with unmanned support, to unmanned formations with humans in command.’ The implication of ‘humans in command’ rather than ‘humans in control’ is not trivial: it describes a model where human intent sets the mission objective and the system executes all tactical decisions autonomously.

In January 2026, the Pentagon launched a $100 million prize challenge to develop voice-controlled, autonomous drone swarming technology — explicitly requiring that the system translate a soldier’s spoken tactical instruction into coordinated manoeuvres for a swarm of autonomous drones. The human input is a voice command. Every subsequent decision is machine-made. The question of where the ‘human in the loop’ sits in that architecture is, at minimum, contested.

VIII. The Hidden Human Cost

NDAA Section 737 and the Moral Injury of Remote Warfare

Embedded in the FY 2026 National Defense Authorization Act, almost unnoticed in the noise around $25 billion munitions rebuilds and Collaborative Combat Aircraft production decisions, is Section 737: a mandate for the Department of Defense to study the psychological effects of UAS operations on military and civilian personnel who operate drones or analyse combat imagery.

The provision calls for an assessment of the prevalence of post-traumatic stress disorder, depression, anxiety, burnout, and moral injury among troops who directly or indirectly operate unmanned aircraft systems — a population that, as the Army’s commitment to one million drones expands, will number in the tens of thousands. Crucially, the study’s scope includes personnel who analyse combat imagery or make targeting assessments — the human nodes in the kill chain who review AI-generated targeting recommendations before approval.

The significance of Section 737 is not merely the welfare of individual operators. It is a formal congressional recognition that the psychological architecture of drone warfare is qualitatively different from any previous form of combat — and that the transition from deliberate, human-paced killing to high-tempo, autonomous-assisted operations creates new forms of moral injury that the military has not yet mapped, let alone addressed.

Research published in the Journal of Mental Health & Clinical Psychology in 2023 found that the remote nature of drone warfare created unique psychological issues, particularly around moral injury: the experience of participating in, witnessing, or failing to prevent events that violate one’s own moral code. As AI systems absorb more of the targeting process, a new form of that injury may emerge — the experience of reviewing and approving AI-generated kill lists that move faster than moral deliberation permits. The ‘human-in-the-loop’ who rubber-stamps 50 AI targeting recommendations per shift is not the same psychologically as the pilot who decided each engagement individually.

Source: DefenseScoop, December 10, 2025; Holland & Knight FY2026 NDAA Analysis; Task & Purpose, December 2025; Journal of Mental Health & Clinical Psychology, 2023

Key Takeaways: The Autonomy Spectrum in 2026

The autonomy spectrum is not moving in one direction — it is fracturing along multiple axes simultaneously. At the tactical level, GPS-denied edge AI is pushing lethal decisions onto the platform itself, away from human operators. At the operational level, agentic AI systems like Maven are absorbing the intelligence and targeting functions that once required human analysts. At the doctrinal level, the mid-flight AI software swap on the YFQ-44A demonstrates that combat autonomy is already being treated as an upgradeable service layer rather than a fixed platform capability.

Against this backdrop, the international governance framework remains rooted in concepts — human-in-the-loop, meaningful human control, appropriate levels of human judgment — that were designed for a slower, simpler version of automation. The Anthropic-Pentagon dispute of February 2026 is not an anomaly; it is a preview of the institutional conflicts that will define AI governance in defence for the next decade. And Section 737 of the FY 2026 NDAA signals that Congress, at least, understands that the human cost of this transformation is not confined to the battlefield.

Key Sources & Expert References

• DefenseScoop: ‘DOD Raises Palantir’s Maven Contract to $1B+,’ May 2025 — defensescoop.com

• DefenseScoop: ‘FY26 NDAA: Psychological Study on Drone Operators,’ December 10, 2025 — defensescoop.com

• Breaking Defense: ‘NATO Picks Palantir’s Maven AI for Military Planning,’ April 14, 2025 — breakingdefense.com

• Air & Space Forces Magazine: ‘Anduril CCA Switches AI Pilots Midflight,’ February 28, 2026 — airandspaceforces.com

• The Aviationist: ‘YFQ-44A CCA Tests Shield AI’s Hivemind and Anduril’s Lattice AIs During Single Flight,’ March 3, 2026 — theaviationist.com

• Anduril Industries: ‘YFQ-44A Flies with Mission Autonomy Software from Anduril & Shield AI,’ February 26, 2026 — anduril.com

• Shield AI: Hivemind Architecture and DT25 Demonstration Documentation, August 2025 — shield.ai

• TechCrunch: ‘Anthropic vs. the Pentagon: What’s Actually at Stake?’ February 27, 2026 — techcrunch.com

• Holland & Knight LLP: ‘FY2026 NDAA: A Comprehensive Analysis,’ December 2025 — hklaw.com

• Congressional Research Service: ‘Defense Primer: U.S. Policy on Lethal Autonomous Weapon Systems’ (IF11150) — congress.gov

• War on the Rocks: ‘Autonomous Weapon Systems: No Human-in-the-Loop Required, and Other Myths Dispelled,’ May 2025 — warontherocks.com

• UN General Assembly Resolution A/RES/79/L.69 on LAWS Governance, December 2024 — un.org

• Task & Purpose: ‘Congress Wants to Screen Drone Operators for PTSD, Depression,’ December 2025 — taskandpurpose.com

• Dronelife: ‘Auterion Demonstrates First Multi-Manufacturer Combat Drone Swarm,’ December 2025 — dronelife.com

• Wikipedia / Project Maven (updated March 2026) — Cross-referenced against primary sources