Matched MCP and CLI interfaces drive agents onto different trajectories in 94% of paired runs

Same tools, same runtime, different wire format, and the agent still walks a different path.

Shi, Pang, and Li posted arXiv:2610.34298 on 25 October 2026 with a controlled comparison that treats MCP versus CLI as part of the decision environment rather than as a cosmetic wrapper. The result that travels is blunt: when executable capabilities are matched, interface realisation still splits trajectories.

MICA locks the contract; DART repairs how the model says the action

Matched Interface Comparison and Analysis (MICA) builds MCP and CLI presentations from one shared semantic action contract. A decoder maps either surface response back to the same semantic action; validation, compilation, and environment execution stay shared. The model-facing difference is presentation, response format, and feedback, while the executable tool set stays shared.

Across six models and three interactive environments, those matched interfaces still produce substantial differences in task performance and interaction cost. Effects reach semantic action selection, failure pathways, and recovery, not only parse failures. Matched-state probes localise sensitivity at capability presentation, realisation, and feedback.

Decoupled Action Realization and Targeted Repair (DART) is the intervention. Once a semantic action is validated, DART keeps that commitment and repairs only the interface-specific expression when realisation fails. On the authors' evaluation, DART raises benchmark score in all six evaluated benchmark-interface settings by up to 9.0 points and reduces MCP token cost in all three environments by up to 27.1k tokens per task.

MCP rollouts made the wrapper look free; MICA prices the wrapper

Teams adopting MCP often assume that exposing the same verbs through tools instead of shell commands is a packaging choice. MICA lands as MCP fleets scale and as CLI agents remain the default in many harnesses. The paper's paired-trajectory result, structured divergence in 94.3% of 2,004 episodes, gives operators a numbered reason to A/B the interface, not only the model.

Amplify digests around 27 October 2026 pushed the work into the same week as other MCP reliability papers. The distinct spine is matched capabilities with unmatched behaviour: the interface is inside the policy loop.

arXiv preprint; six models, three environments, 2,004 paired episodes

Paper: arXiv:2610.34298, posted 25 October 2026, authors Zhenning Shi, Liang Pang, and Yongqi Li. Evaluation spans six models and ALFWorld, WebShop, and ScienceWorld. Paired natural trajectories cover 2,004 episodes with 1,889 (94.3%) showing structured divergence. DART results are reported for CLI and MCP on each environment, including Gemini-2.5-Flash Base-DART pairs with score deltas up to +9.0 on ALFWorld MCP and MCP token cuts of 27.1k / 11.2k / 4.3k per task across the three environments.

Treat MICA and DART as a research prototype and analysis framework. Public code release is not confirmed as a pinned production package in the abstract, so reproduction stays paper-first until a repository pin is verified. Read licence terms from that repository if a code drop appears.

Reach for MICA-style A/Bs when MCP and CLI claim the same verbs

Pick this framing when your agent product offers both MCP tools and CLI/shell tools for overlapping operations, or when you are migrating a CLI harness to MCP and need to know whether scores will move for reasons other than capability gaps. Run matched-contract comparisons before declaring the migration behaviour-neutral.

If every tool exists on only one interface and you have no migration or dual-surface plan, full MICA machinery can wait. If recovery after malformed tool calls is already your top incident class, DART's commitment-preserving repair is the nearer design pattern to trial in the controller.

Matched contracts can still hide decoder and feedback mismatches

MICA's claim depends on a truly shared semantic contract and shared downstream runtime. A sloppy port that renames parameters, truncates feedback, or changes validation strictness reintroduces capability mismatch and breaks the comparison. DART helps only after a semantic action is validated; it will not rescue a wrong high-level choice.

Absolute score deltas are tied to the authors' model list, episode budgets, and environment ports. Re-ranking under a different MCP schema dialect or CLI parser needs a fresh run. Trajectory divergence rates are descriptive of this suite; they are not a universal MCP tax.

Paper first for the matched-contract protocol

The arXiv abstract and PDF are the primary record for MICA, the 94.3% divergence figure, and DART's score and token results. Operator takeaway: interface realisation belongs on the evaluation checklist beside model and tool choice.