Sense signals and events
Observe business state and external change, initiating or adapting work when authorized conditions emerge.
Back to Technology
GEA Architecture
Long-Horizon Agent Runtime
Keep work moving toward long-horizon goals and outcomes.
Continuously sense enterprise and external signals, manage goals, plans, and durable state across interactions and workflows, and adapt the next action to intermediate results. Through the Harness, the runtime uses context, skills, and tools so work can pause, resume, and retry while remaining under human oversight and enterprise boundaries.
How It Works
Keep goals continuous across interactions and systems.
The long-horizon runtime extends a conversation into sustained execution: sensing change, initiating work, preserving state, and using the Harness to access context, skills, models, and tools.
Preserve sessions and state
Use events and checkpoints to preserve goals, plans, dependencies, and intermediate outcomes so work can pause and resume across time.
Execute through the Harness
Assemble context for the current step, route models, skills, and tools, and record inputs, outputs, and state changes for each call.
Recovery and escalation
Replan from intermediate outcomes, distinguish safe retry from compensation, and bring people in at low-confidence or high-risk moments.
Technical Artifact · One Complete Run
How one question becomesa traceable enterprise decision.
In a product innovation task, models, context, runtime, and enterprise foundations operate as one sequence—not isolated modules.
Example task / Input
Redesign the opening experience of a chocolate gift box within defined cost and brand constraints.
Task result / Output
A sourced report, an approved direction, and decision memory written back.
Active phase
Frame the task
01020304050607
Initiate
A product innovation question enters
The business objective and constraints become a durable task.
Capabilities in action
Intent understandingDurable task
Validation & Guardrails
Long-running work requires explicit state, failure, and stopping semantics.
Reliable operation goes beyond retry. It depends on where state is checkpointed, whether external actions can be replayed safely, how sandboxes isolate execution, and when work must stop and return to people.
How We Measure
01
Long-task completion
02
Recovery success
03
Human intervention rate
04
Validation pass rate
Boundaries & Guardrails
Validated state checkpoints
Checkpoint goals, inputs, intermediate results, and external-system state at consequential steps, then validate where execution should resume.
Isolated execution and safe recovery
Run tools in controlled workspaces or sandboxes, separating safely replayable actions from external effects requiring compensation.
Authorization, escalation, and stops
Set thresholds for low confidence, value judgments, and consequential actions so work pauses, requests approval, or stops when required.
Connected Technology
Every capability,working as one system.
Explore the technologies that provide enterprise context, task evaluation, and governance for long-running work.
01
GEA Architecture
Context System
Help agents understand what is happening in the enterprise—and keep acting on it.
Learn more02Enterprise Agent Foundation
Agent Reliability & Evaluation
Make every run observable and measurable—and improve the next one.
Learn more03Enterprise Agent Foundation
Security & Governance
Keep every action within identity, access, and accountability boundaries.
Learn moreTechnical questions
Understand the mechanism, boundaries, and production requirements.
01
How is a Long-Horizon Agent Runtime different from a fixed workflow?
Fixed workflows suit tasks whose steps and branches can be defined in advance. A long-horizon runtime supports work whose path changes with new evidence, intermediate outcomes, and external events. It preserves goals and state, then adapts the next action within constraints instead of pre-writing every case.
02
Which tasks are a good fit for a Long-Horizon Agent Runtime?
It is best suited to work that runs over time, spans systems, requires repeated judgment, and changes direction as intermediate results arrive. When steps are fully deterministic, short-lived, and context does not evolve, conventional automation is usually simpler.
03
How do runtime state and the Context System differ?
Runtime state tracks where one task currently stands, including its goal, plan, dependencies, and intermediate outcomes. The Context System organizes enterprise knowledge, rules, relationships, and decision memory that can be reused across tasks. The runtime reads context to advance work, then writes validated outcomes back to complete the learning loop.
Ready when you are