Skip to content

Execution Model

Developer Reference

This page covers internal implementation details. It is not included in the User Guide.

The Platform Framework routes every request through a deterministic sequence of phases. Three execution flows. Three agent formulas. All stateless.


Three execution flows

FlowTriggerOutput
StandardSingle-domain promptFinding, Report, Triage — one skill, one LLM reasoning pass
BroadMulti-domain or estate-wide questionSynthesised multi-domain report — parallel skill execution
PlaybookWrite-capable operation or explicit triggerBundle → Run → Evidence — human approval always required

Phase routing

Every request enters the Gateway and is classified through a deterministic sequence of phases:

Target assignment at Phase 1:

Target codeRoutes toWhen
HAnalysis AgentQuestions, analysis requests, read-only intent
PPlaybook AgentWrite operations, remediation triggers
RAutoresolverAutomated resolution rules (no LLM call)

The three agent formulas

Each flow maps to a different formula. The formula is deterministic — given the same inputs, the same phases execute in the same order.

Formula 1 — Standard (single domain)

Formula 2 — Broad (multi-domain)

Input:   "Are any of my overprovisioned instances also security risks?"
Phase 1: classify → target = H, mode = broad
Phase 2: resolve skills → [finops.idle_detection, security.detect_public_ingress]
Phase 3: assemble combined context for both skills (parallel CE calls)
Phase 4: Analysis Agent executes both skill tool plans in parallel
         Synthesises joint response correlating cost + security signals
Phase 5: UI Agent renders synthesised multi-domain report
Output:  streamed cross-domain analysis with correlated findings

Formula 3 — Playbook


Stateless guarantee

The Platform Framework has no session state. Each request is self-contained:

  • All conversation history is sent by the client in the prompt field
  • All estate context is assembled client-side and sent in the context field
  • Backend agents retain nothing between requests
  • Scaling out backend agents requires zero session synchronisation

State lives in two places:

  • Client (Tauri app): EstateView, conversation history, artifacts, evidence — in local databases
  • Context Engine (server): registered skills, agents, tools, guardrails, templates, domain knowledge — in collections

Next steps

Escher — Agentic CloudOps by Tessell