ข้ามไปยังเนื้อหา
CephalonEngine

Cephalon.Agentics

เนื้อหานี้ยังไม่ได้แปลเป็นภาษาไทย แสดงเป็นภาษาอังกฤษแทน

Maturity: M3 · Ownership: mixed: application-managed + cephalon-managed · Family: agentics · See audit, matrix.

Cephalon.Agentics is the baseline technology pack for agentic workloads.

What it owns

Section titled “What it owns”
  • agentic runtime options
  • module and registration entry points for the AgenticWorkloads technology
  • tool descriptors, registries, and catalogs
  • orchestration-link validation for tool descriptors that point back to capabilities, execution graphs, or hosted executions
  • Cephalon-managed agent-tool dispatch through registered executors
  • bounded process-local retry for executor failures when explicitly configured
  • process-local duplicate-completed run suppression when explicitly configured
  • run-state reporting for tool executions, including retry-scheduled, approval-required, duplicate-completed, and terminal-failure posture
  • policy and observer hooks for approval, denial, audit, and operational projection
  • runtime-surface contribution for introspection

Main surfaces

Section titled “Main surfaces”
  • Configuration/AgenticRuntimeOptions.cs
  • Modules/AgenticsModule.cs
  • Registration/AgenticEngineBuilderExtensions.cs
  • Services/AgentToolDescriptor.cs
  • Services/AgentToolRegistry.cs
  • Services/AgentToolCatalog.cs
  • Services/IAgentToolContributor.cs
  • Services/IAgentToolCatalog.cs
  • Services/AgentToolDispatcher.cs
  • Services/IAgentToolExecutor.cs
  • Services/IAgentToolExecutionPolicy.cs
  • Services/IAgentToolExecutionObserver.cs
  • Services/IAgentToolRunReporter.cs
  • Services/AgentToolExecutionContext.cs
  • Services/AgentToolExecutionReport.cs
  • Services/AgenticsRuntimeSurfaceContributor.cs

Shared read/action contracts:

  • Cephalon.Abstractions.Agentics/AgentToolExecutionOutcomes.cs
  • Cephalon.Abstractions.Agentics/AgentToolExecutionRequest.cs
  • Cephalon.Abstractions.Agentics/AgentToolExecutionResult.cs
  • Cephalon.Abstractions.Agentics/AgentToolRunState.cs
  • Cephalon.Abstractions.Agentics/IAgentToolDispatcher.cs
  • Cephalon.Abstractions.Agentics/IAgentToolRunCatalog.cs

Source structure

Section titled “Source structure”
  • Configuration
  • Modules
  • Registration
  • Services

How it fits

Section titled “How it fits”

This pack is the reference pattern for future AI or agent runtime behavior in Cephalon. Modules can contribute tools without forcing the host to own the entire catalog, and those tools can stay grounded in the existing runtime model by linking back to published capability keys, execution graphs, and hosted executions.

When AgenticRuntimeOptions.EnableExecution is enabled, the pack also owns a narrow managed execution lane:

  • the abstraction-level IAgentToolDispatcher contract resolves the selected AgentToolDescriptor, reports a started observation, evaluates IAgentToolExecutionPolicy hooks, invokes the matching IAgentToolExecutor, and reports the final outcome through the Cephalon.Agentics implementation.
  • when AgenticRuntimeOptions.ExecutionMaxAttempts is greater than 1, the dispatcher retries failed executor attempts inside the same process, reports retry-scheduled observations, and publishes retryPolicy = bounded-in-process, max-attempt, delay, durability, scope, and next-attempt metadata without claiming durable retry queues.
  • when AgenticRuntimeOptions.EnableExecutionIdempotency is enabled, the dispatcher suppresses duplicate toolId + runId executions that already completed successfully inside the current process, reports the duplicate as skipped, and publishes idempotencyPolicy = completed-run, idempotencyKey = tool-run, retention, durability, scope, and idempotencyOutcome = duplicate-skipped metadata without claiming a durable inbox, broker deduplication, cross-node exactly-once delivery, or provider AI orchestration.
  • IAgentToolRunCatalog exposes the latest run-state truth for each tool run through Cephalon.Abstractions.Agentics, including outcome counts, retry-scheduled counts, skipped counts, actor/correlation details, approval-required posture, retry-pending posture, duplicate-completed posture, terminal-failure posture, terminal-state posture, and the latest operator metadata.
  • IAgentToolExecutionObserver receives every report after it is recorded so modules can attach audit, telemetry, or projection behavior without replacing the dispatcher.

The resulting operator-facing answer flows through /engine/technology-surfaces, /engine/agent-tool-runs, /engine/agent-tool-runs/retry-pending, /engine/agent-tool-runs/idempotency-duplicates, /engine/agent-tool-runs/approval-required, /engine/agent-tool-runs/terminal-failures, /engine/agent-tool-runs/{runId}, /engine/agent-tool-runs/by-tool/{toolId}, POST /engine/agent-tools/{toolId}/runs, and snapshot.AgentToolRuns. Tool entries now include execution readiness (executionEnabled, executionOwnership, executorConfigured, executorCount) and run-state metadata (runtimeState, runCount, lastOutcome, retryPolicy, retryMaxAttempts, retryDelayMilliseconds, retryScheduledCount, retryPending, idempotencyPolicy, idempotencyKey, idempotencyRetentionMinutes, idempotencyDurability, idempotencyScope, duplicateCompleted, terminalFailure, totalReports, approval/denial counters, and reported.* metadata). A tool without a registered executor remains truthful as awaiting-executor rather than being described as fully managed.

ASP.NET Core hosts can trigger one bounded managed run through POST /engine/agent-tools/{toolId}/runs without referencing Cephalon.Agentics implementation types. The route accepts optional run id, actor id, correlation id, attempt, string arguments, and metadata, adds safe trigger and route metadata, and returns 404 when the active runtime does not expose the dispatcher or the selected tool does not exist.

This is still intentionally not an autonomous agent planner, durable approval workflow, durable retry queue, durable inbox, memory store, distributed scheduler, cross-node exactly-once layer, dead-letter system, or provider-specific AI orchestration layer. Those should land only when a package owns those runtime paths explicitly.

The showcase sample now includes ShowcaseAgenticsModule, which contributes a catalog-inspection tool, a matching executor, an approval policy, and an observer hook so the dispatcher/run-state loop is proven end to end in an adoption-quality host.

Section titled “Related docs”