Orchestrating Autonomous Systems in DDIL Environments: Pre-Authorized Envelopes, Lost-Link Behavior, and Reconciliation
How to design orchestration for autonomous platforms and AI agents when communications are denied, degraded, intermittent, or limited: bounded authority issued in advance, fallback rules per action class, and a reconnect protocol that puts deferred decisions back in front of a person.
Most orchestration designs for autonomous platforms and AI agents assume that when the system needs a person, a person can be reached. In a contested environment that assumption fails first. The Atlantic Council’s August 2026 report AI and autonomy in contested operations states the planning case plainly: United States forces should presume that most future operations will take place where connectivity is poor or nonexistent, under denied, degraded, intermittent, and limited (DDIL) conditions caused by jamming, cyber attack, and anti-satellite weapons.
Our earlier post on scoped authority and approval tiers for AI agents described an orchestration layer that asks a named approver before a consequential action. This post covers what that layer should do when the approver cannot answer: issue authority before the link drops, choose fallback behavior per action class rather than per platform, and treat the reconnect as a review event, not a resume button.
The lost link is a policy problem before it is a networking problem
Mesh radios, store-and-forward messaging, and edge compute all help. None of them answer the design question: what is each system allowed to do on its own, for how long, and what happens to the decisions it could not make?
Existing policy contains the shape of an answer for one class of systems. The Congressional Research Service summary of Department of Defense Directive 3000.09 notes that covered weapon systems must complete engagements within a timeframe and geographic area consistent with commander and operator intentions, and otherwise must terminate the engagement or obtain additional operator input before continuing. A June 2026 presidential memorandum, NSPM-11, directed an update to that directive within 90 days and defined controllability as the ability to monitor the operation and outcomes of a system and take corrective action as needed. Programs should write requirements against whichever version of the directive is current when they do so.
Two points carry over to every autonomous system, armed or not. Authority is bounded in time and space and granted in advance, and when the bound is reached without a person available, the default is to stop or hold. Most orchestrated systems in logistics, reconnaissance, maintenance, and cyber defense sit outside that directive’s scope, so the program office has to write the rule itself.
Pre-authorized envelopes: issue bounded authority before departure
An envelope is the set of actions a platform or agent may take without further approval, together with the conditions under which the set applies. A usable envelope has four parts.
- Action classes. Which approval tiers may proceed autonomously: observe and propose almost always, reversible acts in defined cases, irreversible or external acts only where policy explicitly permits.
- Spatial and temporal bounds. The area and the clock. When either expires, the envelope collapses to its most conservative state regardless of what the planner wants.
- World-state assumptions. The conditions the approval depended on, such as an unchanged no-go list or sensor data no older than a stated age. If a condition fails, the actions that depended on it are withdrawn.
- A signed record. Who approved the envelope, when, and what evidence they saw. This is what the reviewer reads after reconnect.
The envelope should be enforced by code on the platform, not described in a prompt or a mission brief. A monitor that checks each proposed action against it is the simplex structure from our post on runtime assurance for autonomous systems. Under DDIL that monitor is the only supervisor present, so it must run locally without reach-back.
Fallback behavior per action class, not per platform
A single lost-link mode for a whole platform is too blunt. A reconnaissance vehicle that loses its link should probably continue its surveyed route, but it should not start a task that pulls coverage away from a sector another vehicle depends on. The fallback rule belongs to the action. The table below is illustrative; the cells are design choices, not doctrine.
| Action class | Link up | Link intermittent | Link denied beyond timeout |
|---|---|---|---|
| Observe (sense, log, summarize) | Proceed, stream | Proceed, buffer | Proceed, buffer locally |
| Propose (replan, recommend) | Proceed, send for review | Proceed, queue for review | Proceed, queue; hold any dependent act |
| Act, reversible, inside envelope | Proceed, log | Proceed, log | Proceed only if world-state assumptions still hold |
| Act, reversible, outside envelope | Request approval | Request, retry until timeout | Defer; hold at last approved state |
| Act, irreversible or external | Request approval with evidence | Request, retry until timeout | Defer; execute pre-approved safe behavior (hold, return, loiter) |
Three details make this work. The timeout that moves a system from intermittent to denied should be set per action, since a reversible action can wait minutes while a time-critical sensing task cannot. A deferred request must be recorded in the same format as a live approval request, so the reviewer later sees the same evidence. And each platform needs a deterministic rule for which peer coordinates when a group is cut off from the operator but not from each other; otherwise two halves of a partitioned group can each conclude they are in charge and reassign the same task. The May 2026 guidance from CISA and partner agencies, Careful Adoption of Agentic AI Services, lists cascading structural failures among its five risk categories; a partition is the physical form of that risk.
Reconciliation: treat the reconnect as a review, not a resume
When the link returns, the operator’s picture and the platform’s record have diverged. The orchestrator should run a fixed sequence before issuing any new authority.
- Exchange logs with sequence numbers. The platform sends every action, deferred request, and monitor decision since the last acknowledged entry, and reports any gaps.
- Replay against the envelope. Each autonomous action is re-checked against the envelope and the world-state assumptions as they stood at the time. An action that fell outside the envelope should not exist; if one does, it is a fault and goes to the top of the review.
- Surface deferred decisions to a person. Every queued request is presented in order with the evidence captured at the time. Some will be moot. The person closes each one explicitly.
- Rebase the plan. Only then does the orchestrator accept a new plan and issue a new envelope. Credentials tied to the old envelope expire with it.
This sequence is also where test and evaluation gets its evidence. The Atlantic Council report recommends testing and red-teaming AI against battlefield conditions it never encountered in training. For orchestrated systems, the specific test is to cut the link at every decision point in a scenario, not only at the scripted moment, and to check that the log reconciles cleanly each time.
An illustrative example: four surface vessels and a forty-minute outage
Four uncrewed surface vessels survey a channel ahead of a convoy. Their envelope, approved that morning, covers the channel and a two-mile margin for six hours, permits route adjustments within the surveyed area, and requires approval for any contact investigation outside the margin. Its world-state assumptions include sensor data no older than fifteen minutes. All numbers here are illustrative.
Forty minutes in, jamming severs the operator link. Three vessels still see each other; one is cut off entirely. The three elect a coordinator by the pre-set rule (lowest identifier with a healthy monitor) and continue the survey, buffering observations. The planner proposes investigating a contact outside the margin; the local monitor defers it as an act outside the envelope and holds the vessel on its surveyed track. The isolated vessel continues until its freshness assumption fails at fifteen minutes, then holds position, its pre-approved safe behavior.
When the link returns, the orchestrator pulls four logs, finds the deferred investigation and the hold event, and shows both to the operator with the evidence captured at the time. The operator declines the investigation (the convoy has rerouted), releases the held vessel, and issues a new envelope. Nothing here required a smarter planner, only authority written down in advance and a reconnect that put the deferred choices in front of a person.
Why this matters now
The January 2026 Department of War artificial intelligence strategy, as reported by Breaking Defense, names an Agent Network project aimed at semi-autonomous agents for battle management and decision support. Analysts at CSIS note that today’s uncrewed systems depend on a continuous operator link and most become useless once it is severed. More autonomy is coming where the link is least reliable. The envelope, the per-action fallback, and the reconcile step are how a program keeps controllability, in the NSPM-11 sense, while the person is out of the loop.
How this connects to our work
MissionGuard checks each course of action against declared mission contracts, including contingency response and information freshness, and records a structured fault card when one is violated. Those contracts are a natural place to encode an envelope’s world-state assumptions for local enforcement. AEGIS applies the same evidence-chain discipline to coordinated AI agents in cyber inspection and defensive validation under human supervision.
If your program is specifying lost-link behavior for autonomous platforms or AI agents and wants to work through envelopes, fallback rules, or the reconcile sequence your authorizing official will expect to see tested, we would be glad to talk. Contact QuantumWorks.