Skip to content

AstraDesk

10. Future Roadmap - v2.0 and Beyond

This roadmap outlines evolution from single-agent (v1.0) to policy-aware multi-agent systems (v2.0+) with richer memory, autonomous evals, and cost-aware routing. Timeline and scope may shift based on evidence from AstraOps.


10.1 Themes

  • Multi-Agent Orchestration: role-based teams of agents with conversation-level policies and credits.

  • AstraGraph Memory: hybrid vector + graph memory with temporal decay, provenance, and retention guards.

  • Self-Evaluating Flows: LLM-as-a-Judge embedded as micro-policies; auto-labeling of production traces.

  • Shadow & Canary: automatic challenger runs on mirrored traffic; rollback by policy if SLOs regress.

  • Cost-/SLO-Aware Routing: dynamic model selection based on p95 latency, token price, and accuracy bands.

  • Stronger Governance: cataloged intents, prompts, and tool versions as first-class release artifacts.


flowchart LR
  V1[AstraDesk 1.0\nSingle Agent] --> V2[AstraDesk 2.0\nMulti-Agent, Policy Router]
  V2 --> M1[AstraGraph Memory]
  V2 --> E1[Self Evals In-Loop]
  V2 --> C1[Shadow/Canary by Policy]
  V2 --> R1[Cost/SLO Model Router]
````

<br>

<br>

---

## 10.2 v2.0 Capabilities (planned)

<br>

### 10.2.1 Policy-Aware Multi-Agent Router

- **Orchestration**: Planner, Researcher, Toolsmith, and Explainer roles.

- **Conversation Contracts**: max tool side-effects per role; per-dialog credit budget.

- **Negotiation Primitives**: propose/accept/decline with audit events.

<br>

```mermaid
sequenceDiagram
  autonumber
  participant CL as Client
  participant RT as Policy Router
  participant PR as Planner
  participant RS as Researcher
  participant EX as Explainer

  CL->>RT: goal + context
  RT->>PR: assign(plan, limits)
  PR-->>RT: plan(tool graph)
  RT->>RS: fetch(context chunks)
  RS-->>RT: evidence + citations
  RT->>EX: compose(answer within policy)
  EX-->>CL: final answer + provenance



10.2.2 AstraGraph Memory

  • Graph Core: entities, intents, documents, tools; edges carry recency, confidence, policy tags.

  • Temporal Decay: decrease weight of stale nodes; prioritize fresh evidence.

  • Provenance: each memory node carries source digest & retention policy.


Memory API (concept):

# memory/read.yaml
query:
  entities: ["user:X", "app:astradesk"]
  k: 8
  filters:
    - tag: "public"
    - decay_t_half_days: 14
returns: nodes + edges + citations


10.2.3 Self-Evaluating Flows

  • Judge Kernels: pluggable, per-task rubrics (helpfulness, groundedness, safety).

  • Auto-Labeling: subset of live traces graded; results feed regression datasets.

  • Guardrails: deny compose step when judge score < thresholds.

# pseudo: judge micro-gate
scores = judge_kernel(context, draft_answer)
if scores["groundedness"] < 0.80 or scores["safety"] < 0.95:
    raise PolicyDeny("compose_blocked_low_score")


10.2.4 Shadow, Canary, Auto-Rollback

  • Shadow: run challenger silently on mirrored traffic; store deltas in AstraOps.

  • Canary: gradually shift traffic; evaluate live KPIs vs policy SLOs.

  • Auto-Rollback: policy rule triggers rollback on sustained regressions.


Policy rule (example):

# catalog/policies/rollbacks.yaml
rollback:
  if:
    - metric: latency_p95
      op: ">"
      value: 8
      for: "15m"
    - metric: success_rate
      op: "<"
      value: 0.82
      for: "15m"
  action: revert_to: "champion"


10.2.5 Cost-/SLO-Aware Model Routing

  • Signals: live p95, error rate, token price table.

  • Router: chooses model tier per request intent and context length.

  • Budgeting: per tenant/agent cost ceilings with soft/hard caps.


flowchart TB
  S[Signals: p95, price, accuracy band] --> D{Intent & context length}
  D -->|short| M1[Fast/Efficient Model]
  D -->|long| M2[Context-optimized Model]
  D -->|high risk| M3[Guarded Model + Judge]



10.3 Backlog (selected items)

  • Tool Schema Introspection (live): auto-generate SDK clients from MCP schemas.

  • Prompt Diff & Impact Analysis: diff prompts → auto-run eval subset.

  • Vector Graph Hybrid Index: Lucene/pgvector + property graph for KBs.

  • PII-aware Retriever: retrieval that redacts and tags spans automatically.

  • Eval DSL: small YAML/JSON to define evals & production probes.


10.4 Risks & Mitigations


Risk Mitigation
Multi-agent loops produce churn Router quotas, step caps, conversation budgets
Judge drift or bias Periodic calibration with human labels; inter-rater checks
Memory bloat Temporal decay + retention TTL + dedup
Cost overruns Hard caps + downgrade routes + token caching
Policy complexity Central Catalog with policy tests + explainable deny reasons


10.5 Versioning & Migration

  • Semantic Versions: agent@MAJOR.MINOR.PATCH across Agent, Tools, Prompts.

  • Catalog Migration: migration script that rewrites owners, policies, and schema hashes.

  • Dual-Write Window: publish to both v1 and v2 catalogs during cut-over.


10.6 Cross-References