Ilmu Komputer & AI editorial
Open AccessOA2026
Can MCP Clients Decide What to Do After Failure? A Result-Only Actionability Audit
A six-part actionability profile reveals that completed MCP failure results often make failure observable but rarely make concrete recovery or safe replay self-contained.
Rishabh Mehanยท 2026ยท DOI 10.48550/arXiv.2609.00072
The core problem
A client that receives `isError:true` knows that something went wrong. It may still have no machine-readable basis for deciding whether to fix an argument, authenticate, wait, choose another tool, or stop. This paper studies what deterministic software can learn from a completed MCP failure result alone; request arguments, schemas, discovery history, authentication state, transport metadata, host policy, and application state are outside that boundary. The central research question is: to what extent can a client, using only the completed MCP failure result, determine an appropriate recovery action? The authors introduce a six-part actionability profile and apply it with record-level evidence. The study is deliberately narrower than an ecosystem survey or an agent benchmark, focusing on the information content of failure results themselves.
Innovation
In the 21 safely induced failures, typed fields expose failure in 18 cases and a broad policy in 8, yet expose no specific cause, target, executable repair, or replay constraint. Prose often carries more cause and target information, at the price of making semantic interpretation part of the recovery path. The lexical source audit finds the same text-centered pattern. Quantitatively, the actionability profile scores are as follows:
- Failure occurrence: 18/21 (85.7%)
- Broad policy: 8/21 (38.1%)
- Specific cause: 0/21 (0%)
- Target: 0/21 (0%)
- Executable repair: 0/21 (0%)
- Replay constraint: 0/21 (0%)
The average actionability score
out of 6, indicating low overall actionability. The fail-closed prototype demonstrates how a separate experimental control plane could support deterministic branching, but this is not part of the standard MCP result.
A client that receives `isError:true` knows that something went wrong. It may still have no machine-readable basis for deciding whether to fix an argument, authenticate, wait, choose another tool, or stop. This paper studies what deterministic software can learn from a completed MCP failure result alone; request arguments, schemas, discovery history, authentication state, transport metadata, host policy, and application state are outside that boundary. The central research question is: to what extent can a client, using only the completed MCP failure result, determine an appropriate recovery action? The authors introduce a six-part actionability profile and apply it with record-level evidence. The study is deliberately narrower than an ecosystem survey or an agent benchmark, focusing on the information content of failure results themselves.
The authors conducted a small illustrative study of 21 safely induced failures from ten reachable sampled servers. They applied a six-part actionability profile to each failure result, examining whether typed fields and prose expose: (1) failure occurrence, (2) broad policy, (3) specific cause, (4) target, (5) executable repair, and (6) replay constraint. The profile is defined as follows:
Why it matters
The results indicate that completed MCP results often make failure observable, sometimes make a broad response possible, and rarely make concrete recovery or safe replay self-contained in this sample. This has significant implications for the design of MCP clients and servers. Clients cannot rely solely on the failure result to determine specific recovery actions; they must either incorporate external context (e.g., request arguments, schemas, authentication state) or rely on semantic interpretation of prose, which introduces nondeterminism. The authors propose a fail-closed prototype as a potential solution, where a separate control plane provides additional information for deterministic branching. However, this approach requires architectural changes and may not be feasible in all deployments. The study is limited by its small sample size (21 failures from 10 servers) and the artificial induction of failures, which may not represent all real-world scenarios. Future work should expand the sample and explore the integration of control planes in production systems. The findings suggest that the current MCP specification may need to be extended to include more structured failure information to enable automated recovery.
Who should read this
CS practitioners and researchers
Opening member contentโฆ