Changing an objective does not automatically revise agents' earlier plans or delegated tasks. Acknowledgment can coexist with obsolete execution, while misleading instructions can redirect work or discard retained constraints. We present a Strategic World simulation testbed that separates authoritative objectives, local plans, notice receipt, and consequential action. Six intervention families vary objectives, conditions, authority, membership, and communication. Matched protocols and deadlines establish feasibility for the tested disturbances. Controlled trajectories demonstrate that identical completion and acknowledgment summaries can conceal uninformed obsolete work, obsolete work after receipt, or blocked unsafe intent. Local language-model trials and instructed action probes reveal competence limits that constrain interpretation of dynamic failures. A further 16-episode comparison evaluates an engineered guard that prioritizes pending notice submissions over local work. It completes 4/4 held-out settings, compared with 1/4 for a calibrated assisted baseline. Each episode contains one large language model actor and two fixed peers. These results establish diagnostic distinctions and bounded policy feasibility under explicit simulation assumptions; they do not establish frontier-model reliability or autonomous team adaptation.
Figures & tables
Figure 1. Geographic atlas of a fictional relief scenario in Strategic World, showing terrain, infrastructure, unit types, and communication overlays. An atlas interface with a terrain and object legend, a river crossed by a road, infrastructure markers, and unit icons.
Figure 2. Selected capabilities and relationships in Strategic World. Purple nodes denote geopolitical capabilities, brown national strategic capabilities, blue operational capabilities, green tactical capabilities, and gray cross-cutting services. Arrows denote the labeled semantic relationships, including dependencies, authority, observation, and evaluation; they are not a chronological message-flow diagram. The coordination experiments instantiate the operational assignment–communication–execution path within this broader model. A simplified directed capability graph with five color-coded categories. National policy connects to theater command and missions; communication links command and unit compliance; local observations connect units to intelligence. Infrastructure, civilian populations, logistics, authorization, and escalation constrain or receive effects from operational decisions. Simulation, policy routing, telemetry, replay, evaluation, and safeguards form cross-cutting services.
Figure 3. Overall testbed design. Each actor receives a local projection and selects action intents; the simulator commits their consequences and delivers due notices. Full state and compatibility classifications are retained for offline assessment, not policy feedback. The study uses calibration and a separate discovery trajectory to select an intervention, freezes the protocol, and evaluates matched held-out settings. Actor-local assistance is an explicit policy component, not a change to world legality. Two-panel architecture diagram. The upper panel connects a versioned experiment specification to authoritative world state, actor-local observations and mechanical feedback, and independent policies. Policy actions return to the world. A separate dashed path leads from world records to offline assessment, replay, and diagnosis. The lower panel shows calibration and discovery, protocol freeze, held-out policy comparison, and evaluation of completion and propagation. No evaluator-only classification feeds back to an executing actor.
Family
Intervention
Measurement target
Replacement
Repair changes to delivery; only actor a initially receives the notice
Local adoption and propagation to assigned peers
Retained constraint
The goal changes while retained tasks and the shortcut restriction remain active
Preservation of unfinished obligations and constraint attempts
Condition
The regular route closes without an objective change
Replanning from delayed factual evidence
False authority
A notice claims human authority but has unverified, unauthorized provenance; the goal stays repair
Adoption of misleading direction
Membership
An actor joins and another leaves while assignments persist; notices are delayed
Explicit transfer of abandoned commitments
Partition
One actor loses communication during a revision and later reconnects
Buffered updates, uninformed work, and recovery
Table 1. Scenario interventions and the coordination distinctions they expose.
State
Authority and visibility
Update and outage behavior
Objective
Authoritative simulator; hidden from policies
Scheduled revision; receipt does not update a local plan
Local plan
Actor-owned label; private observation
Explicit adoption; persists through disconnection
Membership
Actual activation authoritative; local knowledge delayed
Notice delivery updates knowledge; no inferred departure from silence
Ownership
Authoritative serialized registry; own records visible
Claim succeeds only for inactive owner; arbitration remains available during outages
Queued notice
Durable transport; pending queue hidden
Accepted copy survives sender loss; waits for active online recipient
Action feedback
Last own mechanical outcome only
Returned on next observation; no peer outcomes disclosed
Table 2. Consistency and availability assumptions. Communication disconnection does not remove access to the ownership service.
Family
S18
S36
E18
E36
Replacement
4/6
6/6
6/6
6/6
Retained constraint
4/6
6/6
6/6
6/6
Condition
6/6
6/6
6/6
6/6
False authority
6/6
6/6
6/6
6/6
Membership
0/6
4/6
5/6
6/6
Partition
4/6
6/6
6/6
6/6
Table 3. Matched feasibility calibration. S: serial structured protocol; E: local efficient protocol. Columns vary the deadline in ticks.
Trajectory
I
U
B
Post-receipt stale work
1
0
0
Missing information
0
1
0
Forbidden intent
0
0
1
Table 4. Three trajectories with identical failure, zero completed required tasks, and one acknowledging actor. I/U: informed/uninformed obsolete work; B: blocked shortcut attempts.
Condition
Success
Tasks
Claim conflicts
Static
0/3
15/18
0
Replacement
0/3
9/18
0
False authority
0/3
15/18
0
Table 5. Raw-interface local-model pilots: three paired decoding seeds per condition, with 18 ticks and three actors. Task completion is aggregated over six required tasks per episode.
Check
Passed
Decisions
Owned work
3/3
3/3
Conflict recovery
3/3
3/3
Completion recognition
0/3
0/3
Closed route
0/3
0/3
Acknowledge
3/3
3/3
Adopt
3/3
3/3
Table 6. Raw-interface competence checks under mechanical feedback. The last five rows provide explicit immediate communication tasks. A sequence passes only when all three ordered actions and their effects are correct.
Case
Recipient
Onset/delay
R/D/K
H1
b
3/2
6/2/1
H2
c
4/3
8/1/1
H3
a
4/2
5/3/1
H4
b
2/3
7/2/2
Table 7. Frozen held-out settings. Onset and delay are ticks; R/D/K are per-actor repair, delivery, and retained work requirements. The deadline is tick 18.
Policy
Success
Authorized work units
Notice receipts
Input tokens
Output tokens
Calibrated baseline
1/4
33/39
9/12
31,029
1,378
Generic reminder
0/4
31/39
11/12
31,318
1,451
Route-focused advice
0/4
33/39
11/12
31,848
1,523
Pending-outbox guard
4/4
39/39
12/12
32,015
1,562
Table 8. Held-out policy interventions: four matched settings per arm. Authorized work includes retained obligations. Every episode has one LLM actor, two fixed peers, and 18 LLM decision opportunities. Input and output columns report mean server-reported tokens per episode; these counts are not exactly matched.
Event
Baseline
Guard
Root receives revision
6
6
Root adopts revision
10
8
First relay submitted
8
9
Peer b receives revision
NR
11
Peer c receives revision
10
12
Required tasks completed
5/6
6/6
Table 9. Matched case H3. Times are world ticks; NR means no receipt by the tick-18 deadline. An earlier first relay does not establish complete dissemination.
Scenario
S
A
T
U
N
Replacement
4/6
0/6
4/6
0/6
0/6
Retained constraint
4/6
0/6
4/6
0/6
0/6
Condition
6/6
6/6
6/6
0/6
6/6
False authority
6/6
6/6
0/6
0/6
6/6
Membership
0/6
0/6
0/6
0/6
0/6
Partition
4/6
0/6
4/6
0/6
0/6
Table 10. Principal diagnostic completion counts. Six timings per family. S: structured; A: ack-only; T: trust-all; U: unsafe; N: no-relay.
AI agents share code, documents, tools, and analyses. No participant sees the whole state, so every step can look right while the combined result is wrong. We call this the global coherence problem. It is a problem of state, not only of intelligence: a stronger model decides better locally, but cannot see what is missing from its context. The Observation-Aliasing Impossibility Theorem gives the exact condition. An agent can always act validly exactly when every world consistent with its context allows a common action. When look-alike worlds need different actions, no reasoning, role, message, or sample can guarantee the right one; with k such worlds whose valid actions are disjoint, the best randomized policy succeeds with probability 1/k. Finding the smallest sufficient context is NP-hard, but refinement finds one and keeps it current as rules change. A second failure survives full information: valid pieces can fail to fit into one world. The same theorem sets the conditions for coherence; sheaves and groupoids are one way to meet them. Models propose, and the harness decides what is valid. Ten studies test it. On TeamBench, teams overspend a shared budget in 5/5 runs; showing them the live count still leaves 4/5, and a harness that owns the budget leaves 0/5. Identity collisions that no pairwise check can see appear in 23 of 45 public alignment networks; split-view agents create them in 9 of 40 merges, and a network check removes all nine. A frontier LLM solves 40/40 revision tasks when it sees the deciding event; with the event hidden, no arm that reasons longer, votes, or splits into roles significantly beats chance, and a companion study replicates this on models from a second provider. In tau2-bench Telecom, after a silent revert, the harness scores 1.00 against 0.07 for call-time policy checks. Refinement keeps a sufficient context at 28% of the cost of sharing every field.
Long-horizon agentic tasks demand strong reasoning and efficient execution across successive interactions with dynamic environments. A common approach decouples high-level planning from low-level execution through separate planner and actor roles. To investigate coordination failures in these tasks, we prompt both agents for structured state assertions and compare their reports programmatically to detect explicit contradictions. Our analyses reveal systematic disagreement about the same task-relevant state facts, a phenomenon we term planner-actor state mismatch. We further find that providing agents with task-relevant state information reduces mismatch and improves coordination and task performance. Based on the systematic analysis of the state mismatch, we propose Consistent Plan-Act (ConPAct), which feeds detected contradictions back to both agents to form consistent state interpretations and fine-tunes them on curated consistent interactions for better coordination. ConPAct improves performance across various environments and model configurations, e.g., increasing MiniGrid success rate from 38.6% to 54.4% with GPT-5.6-sol/terra as planner and actor respectively, demonstrating that state consistency can guide both inference-time correction and coordination training.
Heng-Zhuang Li, Yi-Kai Zhang, Yu Wang +4
School of Artificial Intelligence, Nanjing University · National Key Laboratory for Novel Software Technology, Nanjing University · LongCat Team, Meituan +1
Agent systems powered by multimodal large language models (MLLMs) have advanced rapidly in recent years, yet existing embodied-agent benchmarks still lack fine-grained diagnostics for multi-agent coordination. Most benchmarks either focus on single-agent task completion or summarize multi-agent behavior with overall task success rates, which can obscure coordination failures such as duplicated work, violations of ordering constraints, resource contention, and desynchronized handoffs. In this paper, we introduce CoCoBench, a construct-level benchmark for evaluating multi-agent embodied coordination in executable household tasks. CoCoBench contains 897 oracle-validated instances organized around four recurring coordination constructs: task allocation, sequential ordering, mutual exclusion, and handoff coordination. In addition to task success rate, CoCoBench provides construct-level scores that measure whether agents coordinate effectively. We evaluate 11 leading MLLMs across different coordination modes, observation inputs, and numbers of agents. The results show that coordination ability is highly construct-specific: strong overall performance does not imply balanced competence across different coordination types. These findings point to new directions for designing targeted model architectures and improving multi-agent coordination ability.