ZTE Maturity Model
The 4 Levels
L1 — MANUAL (score 12–20)
Every prompt is typed by a human. Every output is reviewed by a human. The AI is a tool, not an agent.
Characteristics: no specs, no evaluator, no automated triggers, no lessons stored.
Advantage: 2–5x productivity gain over no AI.
L2 — ASSISTED (score 21–35)
Human writes the spec and reviews the PR. Agent executes the work autonomously.
Characteristics: specs exist, evaluator may exist, some automation, lessons stored but not injected.
Advantage: 10–50x productivity gain.
L3 — AUTONOMOUS (score 36–48)
System detects work, generates specs from intent, executes, and creates PRs. Human reviews and approves.
Characteristics: evaluator blocking gate, lesson injection, multiple trigger types, session continuity.
Advantage: 50–100x productivity gain.
L4 — ZERO TOUCH (score 49–60)
System detects, specifies, executes, evaluates, reviews, and deploys. Human sets direction only.
Characteristics: all 12 leverage points at 4–5, automated deployment gate, outcome feedback loop.
Advantage: 100x+ productivity gain.
Scoring the 12 Leverage Points (1–5 each)
Be strict. Score 5 only when there is clear code evidence of full implementation.
| # |
Dimension |
Score 5 |
Score 3 |
Score 1 |
| 1 |
Spec Quality |
Every feature has testable acceptance criteria with context and edge cases |
High-level spec only, no acceptance criteria |
No spec or vague README |
| 2 |
Context Precision |
Per-intent context compression, lesson injection, relevant examples always present |
Context provided, some noise or irrelevant content |
Raw files dumped, no filtering |
| 3 |
Model Selection |
Auto-selected by task complexity, cost-optimised routing, fallback logic present |
Deliberate model choice, some routing |
Default model for all tasks |
| 4 |
Tool Availability |
Full tool suite + custom MCP servers for domain needs |
Basic tools only (read/write/bash) |
Text generation only, no tools |
| 5 |
Feedback Loops |
Evaluator critique fed back to generator, lessons stored and injected, multi-turn refinement |
Evaluator exists, feedback stored but partial injection |
No feedback mechanism |
| 6 |
Error Recovery |
Retry with backoff, fallback paths, graceful degradation, error classification |
Some retries, most errors handled |
Crashes on any error |
| 7 |
Session Continuity |
Phase-level checkpoints, auto-resume from exact failure point, state persisted to disk |
Session state saved, manual resume possible |
In-memory only, full restart on failure |
| 8 |
Quality Gating |
Hard blocking gate with minimum score threshold enforced, blocking retry with critique |
Gate exists but advisory only |
No quality gate |
| 9 |
Cost Efficiency |
Zero marginal cost architecture, budget tracking per session visible |
Moderate cost, no optimisation |
High cost per operation, no awareness |
| 10 |
Trigger Automation |
Multiple trigger types wired: webhook + cron + Linear poller + Telegram |
1–2 automated triggers working |
Fully manual, human initiates every run |
| 11 |
Knowledge Retention |
Lessons stored, analysed weekly, injected into future generator prompts |
Lessons stored but not injected or analysed |
Nothing retained between sessions |
| 12 |
Workflow Standardization |
PITER/ADW templates enforced, classifier active, all phases defined, no ad-hoc paths |
Informal workflow, some consistency |
Ad-hoc, varies by session |
1---2name: zte-maturity3description: Zero Touch Engineering maturity model — 4-level assessment with explicit scoring criteria per dimension.4---56# ZTE Maturity Model78## The 4 Levels910### L1 — MANUAL (score 12–20)11Every prompt is typed by a human. Every output is reviewed by a human. The AI is a tool, not an agent.12Characteristics: no specs, no evaluator, no automated triggers, no lessons stored.13Advantage: 2–5x productivity gain over no AI.1415### L2 — ASSISTED (score 21–35)16Human writes the spec and reviews the PR. Agent executes the work autonomously.17Characteristics: specs exist, evaluator may exist, some automation, lessons stored but not injected.18Advantage: 10–50x productivity gain.1920### L3 — AUTONOMOUS (score 36–48)21System detects work, generates specs from intent, executes, and creates PRs. Human reviews and approves.22Characteristics: evaluator blocking gate, lesson injection, multiple trigger types, session continuity.23Advantage: 50–100x productivity gain.2425### L4 — ZERO TOUCH (score 49–60)26System detects, specifies, executes, evaluates, reviews, and deploys. Human sets direction only.27Characteristics: all 12 leverage points at 4–5, automated deployment gate, outcome feedback loop.28Advantage: 100x+ productivity gain.2930## Scoring the 12 Leverage Points (1–5 each)3132Be strict. Score 5 only when there is clear code evidence of full implementation.3334| # | Dimension | Score 5 | Score 3 | Score 1 |35|---|-----------|---------|---------|---------|36| 1 | Spec Quality | Every feature has testable acceptance criteria with context and edge cases | High-level spec only, no acceptance criteria | No spec or vague README |37| 2 | Context Precision | Per-intent context compression, lesson injection, relevant examples always present | Context provided, some noise or irrelevant content | Raw files dumped, no filtering |38| 3 | Model Selection | Auto-selected by task complexity, cost-optimised routing, fallback logic present | Deliberate model choice, some routing | Default model for all tasks |39| 4 | Tool Availability | Full tool suite + custom MCP servers for domain needs | Basic tools only (read/write/bash) | Text generation only, no tools |40| 5 | Feedback Loops | Evaluator critique fed back to generator, lessons stored and injected, multi-turn refinement | Evaluator exists, feedback stored but partial injection | No feedback mechanism |41| 6 | Error Recovery | Retry with backoff, fallback paths, graceful degradation, error classification | Some retries, most errors handled | Crashes on any error |42| 7 | Session Continuity | Phase-level checkpoints, auto-resume from exact failure point, state persisted to disk | Session state saved, manual resume possible | In-memory only, full restart on failure |43| 8 | Quality Gating | Hard blocking gate with minimum score threshold enforced, blocking retry with critique | Gate exists but advisory only | No quality gate |44| 9 | Cost Efficiency | Zero marginal cost architecture, budget tracking per session visible | Moderate cost, no optimisation | High cost per operation, no awareness |45| 10 | Trigger Automation | Multiple trigger types wired: webhook + cron + Linear poller + Telegram | 1–2 automated triggers working | Fully manual, human initiates every run |46| 11 | Knowledge Retention | Lessons stored, analysed weekly, injected into future generator prompts | Lessons stored but not injected or analysed | Nothing retained between sessions |47| 12 | Workflow Standardization | PITER/ADW templates enforced, classifier active, all phases defined, no ad-hoc paths | Informal workflow, some consistency | Ad-hoc, varies by session |
cleanexpo