Technical appendix

The emerging agent-computing stack.

A reference map for the layers around intelligence. Field Runtime selects and combines them according to the work—not allegiance to a single vendor.

FR / REFERENCE V1.0

REFERENCE MAP / V1.0

Models are the processor. Runtimes provide persistence and execution. Harnesses become the application. The surrounding layers make the system connected, observable, governable, and useful.

Agentic layerProjectsTraditional analogyWhat it actually does
01IntelligenceGPT · Claude · Gemini · open modelsCPU / processorReasons and generates actions, but does not provide durable identity, memory, scheduling, permissions, or execution by itself.
02Personal agent runtimeHermes Agent · OpenClawOperating systemMaintains a persistent agent, connects channels and tools, schedules work, manages memory and subagents, and mediates model access.
03Evolutionary agent runtimeExoA self-hosting OS that can rewrite its kernelCan inspect and modify prompts, memory, tools, policies, and code; rebuild, restart, rewind, and preserve lineage.
04Application orchestrationLangGraphSpring · Django · TemporalProvides durable execution, state machines, deterministic and agentic steps, persistence, and human checkpoints.
05Domain harnessSupport · revenue · incident agentsExcel · Salesforce · SAPEncodes a job: procedure, context, tools, approval boundaries, exception handling, tests, and measurable outcomes.
06Capability packageAgent Skillsnpm package · plugin · driverPackages instructions, scripts, references, and templates into portable, version-controlled capabilities.
07Tool and data interfaceMCPUSB-C + device-driver interfaceStandardizes how agents connect to tools, databases, applications, and workflows. It solves connectivity, not business logic.
08Agent–client interfaceACPLanguage Server ProtocolStandardizes communication between coding agents and clients such as editors and workspaces.
09Agent-to-agent networkA2ATCP/IP · HTTP · EDILets independently built agent applications discover capabilities, exchange tasks, and collaborate.
10Persistent memoryLettaVirtual memory + stateful databaseProvides durable state and advanced memory so an agent can retain experience over longer periods.
11Secure executionE2BContainers · VMs · process isolationGives agents isolated computers for generated code and commands without unrestricted host access.
12Observability and evaluationPhoenixDebugger + Datadog + test labRecords traces, evaluates outputs, manages datasets and experiments, and diagnoses agent failures.
13Multiplayer work environmentBuzzSlack + GitHub + Jenkins + multiplayer desktopGives humans, agents, workflows, code events, approvals, search, and signed history a shared operating surface.

Field Runtime is an independent implementation company. Project names identify technologies we can evaluate, deploy, and integrate; no affiliation or endorsement is implied.