Agent Mesh — Later capabilities¶
Target
This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.
Nothing here is promised; each item names its trigger and owner. Extension principles kept from the former roadmap: registry-first (new capability is a published release), contract-stable, governed by default, grounded and observable, provider-agnostic behind the gateway, MCP-extensible tools.
Per service¶
| Service | Later capability | Trigger |
|---|---|---|
AgentRuntimeService |
Replay-based simulation against historical traces (BC-22) | run journal and checkpoints in place; a separate ADR, ADR-0014 makes no replay promise |
AgentRuntimeService |
Warm adapter sessions; worker pools per model profile class | latency evidence from the pilot |
AgentRuntimeService |
Outbound A2A calls as tools of class external; computer-use and browser agents as sandboxed tools | first partner agent; Tools & MCP sandbox broker live |
WorkflowEngineService |
Workflow simulation (dry-run cost and duration); visual authoring in Workflow Studio producing the same DSL; in-flight migration to a newer definition on operator decision; cross-run event triggers | usage accounting estimates available; Studio UX track; Fleet Operations telemetry |
AgentRuntimeWorker |
A second execution adapter passing the same contract suite | ADR-0024 revisit trigger |
Platform-wide (from the former roadmap)¶
| Candidate | Purpose | Owner |
|---|---|---|
| Continuous and A/B evaluation of agent versions; prompt-drift detection; prompt and skill optimization from feedback | quality loop | BC-22 |
| Real-time token and cost governor; model failover orchestration per policy | cost and reliability | BC-05, BC-07 |
| Long-running multi-agent coordinator beyond deterministic workflows | supervisor and sub-agent delegation at scale | BC-06 |
| Agent pack installation from a marketplace; evaluation and usage queries; policy dry-run | distribution and inspection | BC-27, BC-29 |
| Vertical, compliance, skill-library, tool-adapter and model-policy packs | marketplace packs | BC-27 |
| Cost-optimization, prompt-engineering, reliability, anonymized cross-tenant insight, compliance-auditor and knowledge-curation agents | new agent roles | Software Factory pack |
Capability-based agent discovery, agent contracts and delegation, goal-oriented operation and additional runtime adapters remain long-term directions. OpenClaw is not a Factory adapter target (D19).