TRANSMISSION // 2609.15397v2

Why Successful AI Tool Calls Can Still Break Workflows

Source published

A paper maps eight ways agent workflows can go wrong despite individual tool operations succeeding, and finds MCP annotations do not fully express the capabilities needed to guard against them.

AI-generated conceptual illustration for Why Successful AI Tool Calls Can Still Break Workflows
AI-generated conceptual illustration; not documentary evidence.

A tool can complete an action in the outside world without the agent receiving confirmation. If the agent retries, it may duplicate the effect; if it assumes success when nothing happened, a required step may be missing. A paper examines this gap between tool operations and the final state of a workflow.

A payment goes through, but the agent never receives confirmation. Should it try again? If the payment succeeded, a retry may charge twice. If the agent assumes success when no payment occurred, the workflow may finish with a required step missing. The challenge is not necessarily that the tool failed: the effect may have happened while the runtime remains uncertain about it.

Artem Trofimov and Boris Novikov examine this problem at the agent–tool boundary in a workshop paper available as an arXiv preprint. They focus on workflows that affect external systems through tools, such as booking, paying or sending messages. Their questions are what can go wrong under retries, parallel actions and partial failures, and what tool interfaces need to expose to help runtimes prevent those failures.

Track the outside world, not just the call

The paper distinguishes events in the world from what software knows about them. A payment either happened or did not; after a timeout, the agent may not know which. Uncertainty is not itself a failure, but acting without resolving it can make retries, workflow commitments or compensation risky.

The authors introduce an “effect-history” model that records tool attempts, external effects, the runtime’s observations and the workflow’s resolution. They use it to organize eight recurring anomaly patterns into three groups: uncertainty about whether an effect occurred, problems with an effect’s timing or survival within a workflow, and conflicts involving other executions or outside observers.

Examples make the categories concrete. A retry can duplicate a payment. An aborted workflow can leave behind a non-refundable booking. A message sent before approval may remain visible if approval is later denied and an attempted withdrawal does not neutralize it. One committed action might also depend on a temporary reservation that another branch cancels. An outside party may react to an action before it is withdrawn; withdrawing the original effect cannot necessarily undo that reaction.

These are not all ordinary planning mistakes. The paper separates an agent choosing the wrong action from external state becoming inconsistent because the runtime cannot confirm, coordinate or control effects. The catalog is coverage-oriented: the authors give structural reasons for its groupings, but say absolute completeness is unavailable in an open world and formal coverage remains future work.

What tool interfaces tell runtimes

From the catalog, the authors derive capabilities a tool boundary would need to support general safeguards. Examples include authoritative status information, a way to identify retries as the same logical action, descriptions of whether and how an effect can be compensated, staging before an effect becomes visible, and coordination for operations on shared resources. These capabilities do not ensure safety by themselves; runtimes must use them through suitable protocols.

The paper also identifies four limits of relying on black-box tool calls alone. Without an authoritative outcome primitive, a runtime cannot both resolve a workflow and guarantee the paper’s specified safety profiles: waiting avoids choosing among retrying, committing or compensating, but leaves the workflow unresolved. Irreversible, non-commuting actions may need mediation to prevent conflicts. Compensation cannot undo an outside party’s reaction to an already visible action. And a runtime operating above multiple tools generally cannot make their releases atomic without those tools participating.

To examine what a shared interface declares, the authors counted annotations on tools returned by responding remote targets in a snapshot of the official Model Context Protocol (MCP) registry dated July 27, 2026. The census included 98,291 tools; it did not call them. Package- and stdio-only entries were excluded because probing them would require executing them, and unreachable targets did not contribute tools. MCP annotations offer coarse hints such as whether a tool is read-only, destructive, idempotent or open-world. The authors report that none of the capabilities they derive is fully expressible in that annotation vocabulary.

The coverage figures need context. Although 74.0% of measured tools serialized at least one annotation and 61.7% serialized all four, one publisher supplied 45.0% of the tools and included all four fields on each. Excluding that publisher, the shares were 52.8% and 30.3%. Annotations are declarations, not verified behavior: a tool may have safeguards internally that its interface does not reveal.

A framework and census, with limits

This paper presents a model, an analysis of interface capabilities and a registry census; it does not measure how often these problems occur in deployed agents. The census describes responding remote targets in one snapshot, not every MCP tool. Since the tools were not called, it cannot establish whether their annotations match their implementations.

The authors motivate reusable transactional contracts at the tool boundary to make relevant properties explicit across systems. Formal proofs of the stated limits and a cost model are identified as future work. The paper is a workshop contribution and arXiv preprint; those publication details and the stated scope are important context for interpreting its claims.

Automatically matched by shared topic keywords. These links do not imply that the sources agree or establish a consensus.