Extension Roadmap¶
Target Architecture — Final-State Design
The Agent Mesh is designed to grow without re-architecture. New agents, skills, tools, and even new services slot into the same registries, event envelope, task contract, and governance model. This page describes the principles that keep extension safe and the concrete directions the final-state platform is built to support.
Extension Principles¶
- Registry-first — new capability is added by registering an
AgentDefinition/AgentVersionorSkillDefinition/SkillVersion, not by changing the runtime. The Agent Registry and Skill Registry are the extension points. - Contract-stable — extensions honor the canonical task contract and event envelope; new behavior is additive and version-suffixed when breaking.
- Governed by default — every new agent, skill, tool, or model policy is authorized by Governance with least-privilege scope before it can act.
- Grounded and observable — extensions consume context packages and emit full telemetry, so new capability is as traceable as existing capability.
- Provider-agnostic — new model providers integrate behind the
ModelRouterServiceand the Integration Platform, not into agent code. - MCP-extensible tools — new tools are added as MCP adapters in the
ToolAdapterServicewithout runtime changes.
Future Microservices¶
| Candidate service | Purpose |
|---|---|
ConnectSoft.Factory.AgentMesh.AgentMarketplaceService |
Publish, discover, and install third-party agent/skill packs (with the Marketplace). |
ConnectSoft.Factory.AgentMesh.AgentEvaluationService |
Continuous offline/online evaluation and A/B scoring of agent versions. |
ConnectSoft.Factory.AgentMesh.PromptOptimizationService |
Automated prompt/skill tuning driven by Observability feedback. |
ConnectSoft.Factory.AgentMesh.AgentSimulationService |
Replay-based simulation of agent runs against historical traces for regression safety. |
ConnectSoft.Factory.AgentMesh.CostGovernorService |
Real-time token/cost budgeting and throttling across tenants. |
ConnectSoft.Factory.AgentMesh.MultiAgentCoordinatorService |
First-class long-running multi-agent collaboration beyond Control Plane workflows. |
Future Workers¶
| Candidate worker | Trigger | Purpose |
|---|---|---|
AgentEvaluationWorker |
Scheduled / AgentTaskCompleted |
Score outcomes and feed agent version rankings. |
PromptDriftDetectionWorker |
Scheduled | Detect quality regression in skill/prompt outputs. |
ModelFailoverWorker |
ModelInvocationFailed |
Orchestrate cross-provider failover per policy. |
CostBudgetEnforcementWorker |
Telemetry threshold | Throttle or pause executions exceeding tenant budgets. |
AgentPackInstallWorker |
AgentPackPublished |
Validate and register installed marketplace packs. |
Future APIs¶
| Candidate endpoint | Purpose |
|---|---|
POST /agent-packs/install |
Install a marketplace agent/skill pack into a tenant. |
GET /agents/{agentId}/evaluations |
Retrieve evaluation scores and ranking history. |
POST /agent-executions/{executionId}/replay |
Replay an execution for debugging/simulation. |
GET /skills/{skillId}/usage |
Reuse and quality analytics for a skill. |
POST /model-policies/evaluate |
Dry-run a model policy against a candidate task. |
Marketplace Opportunities (Agent Packs)¶
- Vertical agent packs — curated clusters tuned for a domain (fintech, healthcare, e-commerce) packaged as installable agent + skill bundles.
- Compliance packs — agents and skills pre-scoped to a regulatory regime (HIPAA, PCI-DSS, GDPR) with policy templates.
- Skill libraries — reusable, versioned skill bundles (e.g. advanced testing, performance tuning) installable across tenants.
- Tool adapter packs — MCP adapters for third-party systems (issue trackers, cloud providers, data platforms).
- Model policy templates — pre-built provider/cost/residency policies for common tenant profiles.
These build on the Template Library Platform and Marketplace distribution model.
Agent Opportunities¶
- Cost Optimization Agent — continuously tunes model routing and prompt size for cost/quality.
- Prompt Engineering Agent — generates and refines skill prompts from evaluation feedback.
- Agent Reliability Agent — monitors agent health and proposes pool/scope adjustments.
- Cross-Tenant Insight Agent — produces anonymized, governed best-practice patterns (within strict isolation rules).
- Compliance Auditor Agent — audits executions against regulatory policy and flags drift.
- Knowledge Curation Agent — improves the Knowledge Platform corpus from execution outcomes, deepening grounding over time.