Agentic lifecycle analysis
Type: Finding
Status: Active supporting research
Canonical for: Product-facing agent lifecycle stages, maturity and design-system/platform boundaries
Depends on: Microsoft, Atlassian and SAP system profiles plus supporting agent-safety evidence
Scope note: Supporting analysis only. Agent lifecycle UI is not evidence of an agent consuming the design system. Use this material only when testing how lifecycle intent might be represented, retrieved, constrained or evaluated.
No Tier 1 source publishes one complete canonical lifecycle. The following stages are a comparative map derived from multiple sources, not official shared terminology.
| Stage | Microsoft | Atlassian | SAP | Supporting evidence | Status |
|---|---|---|---|---|---|
| Intent / invocation | Built for intent; user-initiated language; example prompts | Rovo agents and AI interaction guidance | Natural-language goals for Joule | HAX efficient invocation | Shared concept |
| Capability framing | First-run purpose, examples, limitations/quality | Rovo distinct identity | Trust/transparency guidance | PAIR mental models | Shared concept |
| Context gathering | Identify files, selections and data used | Agents use Atlassian/connected knowledge | Spaces use role, context and enterprise data | Research stresses scope | Shared concept, UI uneven |
| Planning | Not detailed in preserved DS guidance | Agent/product concept, limited public UI pattern | End-to-end orchestration implied | Agent research/practitioner sources | Weak DS evidence |
| Permission / confirmation | User ownership; no detailed permission taxonomy | Limited public DS evidence | Review/control; generation guardrails | OpenAI confirmations/watch mode; 2026 permissions paper | Strong platform need, weak DS standardization |
| Execution | Structured cards/actions within conversation | Rovo agent actions | Agents execute processes across systems | OpenAI agent system | Product/platform implementation |
| Progress / waiting | Completion/error sounds and lifecycle coherence | Limited public shared pattern | Dynamic workspace/process orchestration | Practitioner long-running-task concern | Weak/experimental |
| Interruption / cancellation | Correction/dismissal principles | Limited evidence | Limited evidence | OpenAI supervision and practitioner patterns | Experimental |
| Approval during execution | General user control | Limited evidence | Review/validate and governance | OpenAI high-impact confirmation | Emerging, risk-dependent |
| Failure | Explicit “when system is wrong”; regenerate/edit/context | AI interaction guidance | Safety/trust guidance | HAX recovery principles | Shared at guidance level |
| Recovery / resume | Revise, regenerate, manually edit without restart | Limited evidence | Limited evidence | Practitioner escalation/undo | Partial |
| Completion | Structured response; completion signals | Agent outcomes | Joule Work delivers composed workspace/outcomes | OpenAI product evidence | Shared concept |
| Audit/history | Not detailed in preserved DS set | Limited evidence | “auditable outcomes” at generation time | Permission/audit research | Important but immature |
| Undo / rollback | Correction of output | Limited evidence | Limited evidence | Carbon content revert; supporting agent patterns | Mature for content, immature for external actions |
Shared concepts
The strongest shared lifecycle is modest: users need an accurate mental model, visible context, iterative control, clear errors and recoverability. This is substantially the HAX lifecycle extended to agents.
Missing or underdeveloped states
- Plan preview and plan change.
- Delegated authority and identity.
- Background execution while the user leaves.
- Waiting on external systems or another person.
- Partial completion and resumability.
- Permission scope that changes by action/resource.
- Action-level audit with the initiating human and executing agent.
- Undo when side effects cross systems.
Design-system implications
Potentially reusable: intent entry, capability disclosure, context/source chips, plan/task containers, progress/status, confirmation, interruption, error/recovery and activity history presentation.
Outside DS enforcement: authentication, authorization, tool permissions, transaction boundaries, logs, rollback mechanisms, prompt-injection defenses and model policy. The DS can make these states legible but cannot make the agent safe by itself.
Maturity
- First-run, correction and error recovery: established guidance.
- Context visibility and structured action responses: emerging implementation.
- Progress/interruption/resume: experimental as shared patterns.
- Fine-grained permissions, audit and rollback: platform-critical but not mature DS practice.