Multi-agent large language model (LLM) systems are rapidly emerging as the dominant architecture for enterprise AI automation, yet production deployments exhibit failure rates between 41% and 86.7%, with nearly 79% of failures originating from specification and coordination issues rather than model capability limitations. This paper identifies Semantic Intent Divergence--the phenomenon whereby cooperating LLM agents develop inconsistent interpretations of shared objectives due to siloed context and absent process models--as a primary yet formally unaddressed root cause of multi-agent failure in enterprise settings. We propose the Semantic Consensus Framework (SCF), a process-aware middleware comprising six components: a Process Context Layer for shared operational semantics, a Semantic Intent Graph for formal intent representation, a Conflict Detection Engine for real-time identification of contradictory, contention-based, and causally invalid intent combinations, a Consensus Resolution Protocol using a policy--authority--temporal hierarchy, a Drift Monitor for detecting gradual semantic divergence, and a Process-Aware Governance Integration layer for organizational policy enforcement. Evaluation across 600 runs spanning three multi-agent frameworks (AutoGen, CrewAI, LangGraph) and four enterprise scenarios demonstrates that SCF is the only approach to achieve 100% workflow completion--compared to 25.1% for the next-best baseline--while detecting 65.2% of semantic conflicts with 27.9% precision and providing complete governance audit trails. The framework is protocol-agnostic and compatible with MCP and A2A communication standards.
Figures & tables
Approach
Semantic
Process-
Conflict
Drift
Governance
Framework
Conflict
Aware
Resolution
Detection
Integration
Agnostic
MCP [ 7 ]
✗
✗
✗
✗
✗
✓
A2A [ 8 ]
✗
✗
Partial
✗
✗
✓
AutoGen [ 4 ]
✗
✗
Ad-hoc
✗
✗
✗
CrewAI [ 5 ]
✗
✗
Role-based
✗
✗
✗
LangGraph [ 6 ]
Partial
✗
Graph-based
✗
✗
✗
Table 1: Research gap analysis across multi-agent coordination dimensions. ✓ = addressed; ✗ = not addressed.
Figure 1: Taxonomy of semantic conflict types in multi-agent LLM systems. Type 1 involves logically contradictory intents on the same entity. Type 2 involves competing resource demands that exceed available capacity. Type 3 involves causal dependencies where one agent’s action invalidates another’s preconditions.
Figure 2: Architecture of the Semantic Consensus Framework. SCF operates as a middleware layer between the orchestration and execution layers, intercepting agent actions and analyzing them for semantic conflicts before permitting execution. Internal arrows show the data flow between the six SCF components.
Figure 3: The Consensus Resolution Protocol’s three-tier hierarchy. Conflicts cascade downward only when a higher tier cannot produce an unambiguous resolution. At any tier, a resolved conflict permits the winning action to proceed to execution.
Approach
Precision
Recall
F1 Score
Conflict Rate
Completion Rate
Ungoverned
—
—
—
25.9%
0.2%
Schema-Only
100.0%
21.4%
35.2%
25.9%
0.8%
Judge-Agent
93.2%
66.3%
77.5%
25.9%
25.1%
SCF-NoPCL
28.7%
65.4%
39.8%
25.9%
100.0%
SCF (Full)
27.9%
65.2%
39.0%
25.9%
100.0%
Table 2: Conflict detection and workflow completion across 600 experimental runs (50 runs × 4 scenarios × 3 frameworks).
Figure 4: Comparative performance across all approaches. While the underlying conflict rate remains constant at 25.9%, only SCF achieves 100% workflow completion. The Judge-Agent’s high detection F1 (77.5%) does not translate to workflow reliability because it operates post-execution.
Scenario
Drift Events
Avg Interactions
Completion Rate
Financial Processing
271
9.2
100.0%
Customer Support
365
8.4
100.0%
Supply Chain
429
10.1
100.0%
Software Development
284
8.8
100.0%
Table 3: Drift Monitor activity across enterprise scenarios (SCF Full, aggregated across 3 frameworks).
Framework
Precision
Recall
F1
Completion
AutoGen v0.4
27.8%
65.8%
38.9%
100.0%
CrewAI v0.76
27.0%
64.7%
38.0%
100.0%
LangGraph v0.2
29.3%
64.9%
40.2%
100.0%
Table 4: SCF (Full) performance by framework configuration (averaged across 4 scenarios).