Skip to main content
Primary sources for partial-failure research | Agentic PrinciplesDirect resilience, workflow, and safety sources reviewed for the scoped-progress study.Agentic Principlestroubleshootingagentic-principlestroubleshootingevaluatorresearcherunderstand-safe-agentic-codinggovern-worktroubleshootresearch

Primary sources for partial-failure research

  • Accessed: 2026-08-25
  • Question: When one capability fails, when should useful work continue?
  • Selection: Primary standards, first-party reliability guidance, and original research that directly addresses degraded operation, failure isolation, workflow dependencies, or agent recovery.

Sources and extracted claims​

NIST SP 800-160 Volume 2 Revision 1​

NIST's December 2021 systems-security engineering guidance defines cyber resilience around the ability to anticipate, withstand, recover from, and adapt to adverse conditions. Its Continue objective is to maximize the duration and viability of essential functions during adversity, and its discussion ties that objective to explicit functional dependencies and critical assets. It also states that resilient systems may operate in a degraded state to carry out mission-essential functions.

  • Source: NIST SP 800-160 Vol. 2 Rev. 1, pp. 1, 11–12, 109–110.
  • Supports: continue selected essential work; model dependencies and resource status.
  • Constrains: continuation is about prioritized essential functions, not maximizing arbitrary activity. The publication is normative engineering guidance, not an agent experiment.

AWS Builders' Library — static stability​

AWS separates control-plane changes from data-plane operation. Its static-stability pattern keeps existing correct operation available when a dependency is impaired, while acknowledging that updates from the failed dependency may not arrive. This provides a close operational analogue for separating blocked change actions from unaffected work.

  • Source: Static stability using Availability Zones, accessed 2026-08-25.
  • Supports: isolate dependency effects; keep already-supported behavior working; prepare the degraded path before the incident.
  • Constrains: cached state can become stale. Static stability supports known existing behavior more strongly than novel actions based on unavailable state.

Google SRE — cascading failures​

Google's SRE guidance frames load shedding and graceful degradation as doing as much useful work as a system can without exhausting itself. It recommends selective rejection by task or priority rather than indiscriminate overload. The same chapter warns that rarely exercised degradation paths and naive retries can create new failure modes and cascading load.

  • Source: Addressing Cascading Failures, especially Load Shedding and Graceful Degradation and Retries, accessed 2026-08-25.
  • Supports: useful work should be selected, bounded, and isolated during impairment.
  • Constrains: degraded behavior needs testing, monitoring, disable controls, and retry budgets; more activity is not necessarily more useful work.

Microsoft Azure Architecture Center — bulkheads​

Microsoft's bulkhead pattern partitions resources so a failed dependency does not consume the capacity needed by unrelated services. The guidance says unaffected services can continue, while identifying cost, performance, management overhead, and over-granular isolation as trade-offs.

  • Source: Bulkhead pattern, updated 2026-03-19, accessed 2026-08-25.
  • Supports: failure scope should be contained to consumers that use the failed dependency.
  • Constrains: isolation has real resource and complexity cost and must follow meaningful boundaries.

Zhu et al. — fault-tolerant workflow scheduling​

Zhu et al. formally model a workflow as a directed acyclic graph whose edges carry task precedence and data dependencies, then evaluate fault-tolerant scheduling and resource provisioning. The paper is not about LLM agents, but it supports using dependency structure rather than run-wide state as the unit of scheduling and recovery analysis.

ToolMaze​

ToolMaze evaluates dynamic path discovery and recovery over DAG-shaped tool tasks with explicit versus implicit and transient versus permanent failures. Its reported results show that tool perturbations degrade agents broadly, with implicit semantic corruption especially difficult and complex paths producing trial-and-error loops.

  • Source: Dongsheng Zhu et al., When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents, arXiv:2606.05806v1, 2026.
  • Supports: recovery is a distinct agent capability; DAG structure and failure type belong in the evaluation; blind retries are an inadequate substitute for replanning.
  • Constrains: this is a 2026 preprint over simulated tools. It does not establish that continuing independent real-world effects is safe.

ReliabilityBench​

ReliabilityBench injects timeouts, rate limits, partial responses, and schema drift into tool-using agent tasks and evaluates outcome correctness across repeated and perturbed runs. It reports material success degradation under stress and treats fault tolerance as separate from happy-path task success.

Review result​

The sources converge on a conditional claim: contain a failure to its actual dependency boundary and preserve selected useful function, but do not invent an untested fallback or operate across missing freshness, authority, observability, or consistency guarantees. The literature challenges both global halt and naive continuation. It supports testing a third policy: dependency-scoped continuation with explicit barriers, verification, and recovery state.