Degraded Mode

The core characteristic is the graceful degradation of non-essential functions. The agent's primary objective remains achievable, but secondary features or optimal performance are sacrificed. For example, an e-commerce support agent might lose access to a real-time inventory API but can still answer general product questions using cached data. This involves:

• Priority-based task shedding: Low-priority tool calls or background processes are suspended.
• Fallback to local logic: The agent switches to deterministic, rule-based decision paths where possible.
• Reduced output fidelity: Responses may become less personalized or detailed.

A system in degraded mode must explicitly signal its status through telemetry. This is not a hidden internal state but a first-class observable condition for operators. Key signals include:

• Health check endpoints returning a 200 OK but with a degraded: true header or body field.
• Dedicated metrics like agent_mode tagged as degraded.
• Structured logs indicating which dependency failed and the activated fallback mechanism. This allows monitoring dashboards and orchestrators (like Kubernetes) to distinguish between a failing pod and one operating in a limited capacity, preventing unnecessary restarts.

The agent maintains a minimal viable capability (MVC). The system design pre-defines which components are critical. For instance:

• A multi-agent coordinator might lose its analytics service but continues routing tasks between worker agents.
• An LLM agent with a failed vector database for RAG might switch to using only its parametric knowledge, albeit with a lower accuracy SLO.
• A robotic control agent might enter a safety-constrained mode, limiting speed but maintaining obstacle avoidance. The boundary between core and non-core functions is defined during the agentic SLO definition phase and is enforced by circuit breakers and feature flags.

Degraded mode is reactive, triggered by the failure of external dependencies or internal resource constraints, not by primary logic errors. Common triggers include:

• External API timeouts or 5xx errors from non-critical services (e.g., a weather service for a logistics agent).
• Resource exhaustion, such as hitting a rate limit on a third-party LLM API, causing a fallback to a smaller, cheaper model.
• High-latency responses from a persistence layer that exceed a defined threshold, prompting the agent to work with stale or in-memory data. These are detected via liveliness probes on downstream services or continuous latency monitoring.

The system does not remain passively degraded. It periodically attempts recovery of the failed component. This involves:

• Retry logic with exponential backoff for failed API calls or database connections.
• Re-evaluation of health checks for the problematic dependency.
• State rehydration from a backup source if the primary state persistence layer is unavailable. Successful recovery triggers an automatic transition back to normal operational mode, logged as a state change. Failed recovery attempts may lead to further degradation or, eventually, a controlled shutdown if the MVC is compromised.

Operating in degraded mode has direct, measurable impacts on agent performance benchmarking and agent cost telemetry.

• Performance: Metrics like task success rate, latency, and output quality will deviate from baseline SLOs. These deviations are expected and should be tracked separately (e.g., p95_latency_degraded).
• Cost: The mode may reduce costs (e.g., using cheaper fallback APIs) or increase them (e.g., due to retry loops). Cost attribution must reflect the degraded context.
• Auditing: The agent behavior auditing trail must record the entry/exit from degraded mode and all decisions made within it, as they may follow different logic paths.

A complete, point-in-time capture of an autonomous agent's internal variables, memory contents, and operational status. This is the fundamental data artifact for understanding an agent's condition at any moment.

• Primary Use: Debugging complex failures, performing forensic analysis, and enabling state rollback.
• Example: Capturing the full conversation history, tool call results, and planning context of a customer service agent before it enters a degraded state due to an API outage.

The periodic process of saving an agent's complete operational state to stable storage. This creates known-good recovery points, forming the backbone of failure recovery and stateful deployment strategies.

• Mechanism: Can be full (save entire state) or incremental (save only changes since last checkpoint).
• Critical for: Ensuring an agent can resume from a recent, valid state after a crash, which is a prerequisite for gracefully entering a degraded mode instead of failing entirely.

The mechanism to revert an agent's internal state to a previous checkpoint or snapshot. This is a key recovery action when an agent encounters an unrecoverable error or makes an undesirable series of decisions.

• Trigger: Often automated based on health checks or anomaly detection.
• Relationship to Degraded Mode: A rollback may be performed after an agent has been operating in a degraded mode to restore it to a last-known healthy state before the degradation began.

A periodic signal emitted by an agent to indicate it is alive and processing. It is a fundamental liveness telemetry signal, not a measure of capability.

• Monitoring Use: A missing heartbeat triggers alerts for agent failure. However, an agent can still emit heartbeats while operating in a degraded mode.
• Key Distinction: Heartbeats confirm the process is running; readiness probes and performance metrics are required to detect degraded capability.

A health check mechanism that determines if an agent has fully initialized its state and dependencies and is ready to accept work. A failed probe typically prevents traffic from being routed to the agent.

• Implementation: Often an HTTP endpoint or command that checks database connections, external API health, and memory state.
• Strategic Use: In a degraded mode scenario, a readiness probe might be configured to return a 200 OK but with a header indicating reduced capability, allowing orchestration systems to make informed routing decisions.

A stable, idle condition where an agent is not actively processing tasks, has completed all pending operations, and is conserving resources. It is a normal, healthy state of inactivity.

• Contrast with Degraded Mode: A quiescent agent is fully capable but idle. An agent in degraded mode is active but impaired. Monitoring must distinguish between low activity (quiescent) and low capability (degraded).
• Operational Value: Transitioning to a quiescent state can be a deliberate strategy to conserve resources during widespread system issues.