The Polytech System
POLYTECH v1.2 is a persistent orchestration service, not a single prompt. It combines task state, specialized agents, shared memory, policy enforcement, approvals, testable modules, observability and rollback metadata.
Four planes
Control plane
API, dashboard, task planner, policy evaluator, approval queue and audit records.
Execution plane
Sandboxed modules and connector workers. No arbitrary code, no production connectors by default.
Memory plane
Relational records for tasks, knowledge, approvals, audit events and snapshots — pgvector-ready.
Operations plane
Health endpoint, container logs, backups, metrics, alerting and recovery runbooks.
Task state machine
queued → running → completed queued → awaiting_approval → queued (after approval) awaiting_approval → rejected (after rejection)
Every transition emits an audit event. Before any consequential side effect, the executor re-checks identity, permission, approval freshness, action hash, budget, idempotency key and the connector allowlist.
System boundary
Autonomous by default
Task planning, draft generation, local checks, memory retrieval and sandbox-only work.
Approval required
Publishing externally, changing production configuration, or high-impact operations.
Never self-authorizing
Agents cannot modify policy, grant themselves permissions, or bypass approvals.
Approvals bind to the payload
Exact action hash, connector, environment, scopes, max spend, expiry, reviewer and policy version. Change the payload and the approval is void.
Agent roles
| Agent | Responsibility | Autonomy |
|---|---|---|
| Core Orchestrator | Decompose objectives, assign bounded agents, validate workflow state | P0–P2 |
| Discovery | Identify available tools and sources; never installs unknown tools | P0 |
| Research | Summarize evidence and label uncertainty | P0–P1 |
| Builder | Create proposed plans, prompts, code and artifacts | P1–P2 sandbox |
| QA | Run checks and record pass/fail results | P2 sandbox |
| Security | Evaluate policy, scope, budget, approvals — enforce-only | Enforce-only |
| Executor | Runs only allowlisted, authorized actions | P0–P2; P3/P4 gated |
| Memory | Save and retrieve validated knowledge and artifacts | P0–P2 |
| Monitor | Track state, failures and rollback metadata | P0–P2 |
Learning pipeline: discover → extract → provenance check → quarantine → sandbox test → evaluate → stage → authorize → publish version → monitor → rollback. External content is untrusted input — it can never change system policy.