Leverage Audit

12 Leverage Points Diagnostic

Score each dimension 1–5. The bottom 3 scores (lowest) represent the highest ROI improvements. Total score determines ZTE level: 12–20 Manual | 21–35 Assisted | 36–48 Autonomous | 49–60 Zero Touch.

Dimensions

1. Spec Quality

Is every feature fully specified with testable acceptance criteria?

• 5 — Acceptance criteria, context, edge cases, and constraints documented for every feature
• 4 — Most features specified, minor gaps
• 3 — High-level spec only, no acceptance criteria
• 2 — Vague brief or README
• 1 — No spec. Prompts are typed ad-hoc.

2. Context Precision

Is the right context surfaced at the right time, compressed and noise-free?

• 5 — Per-intent context compression, lesson injection, relevant repo sections always included, noise excluded
• 4 — Context provided, some irrelevant content included
• 3 — Basic context, no compression or relevance filtering
• 2 — Minimal context, mostly raw files
• 1 — No context management. Everything or nothing is provided.

3. Model Selection

Is the right model chosen for each task tier?

• 5 — Auto-selected by task complexity (worker/architect/evaluator tiers), cost-optimised, fallback logic
• 4 — Deliberate model choice per task, some routing logic
• 3 — Single model for all tasks
• 2 — Default model with no consideration of alternatives
• 1 — Wrong model for the task being performed

4. Tool Availability

Does the agent have every tool it needs without human unblocking?

• 5 — Full tool suite (read/write/bash/git/browser) + custom MCP servers for domain needs
• 4 — Most tools available, minor gaps requiring occasional human input
• 3 — Basic tools only, no domain-specific tooling
• 2 — Limited tools, frequent human unblocking needed
• 1 — No tools. Pure text generation.

5. Feedback Loops

Does output quality improve automatically over time without human intervention?

• 5 — Evaluator critique fed back to generator, lessons stored and injected, multi-turn refinement (2–3 cycles)
• 4 — Evaluator exists, feedback stored, partial injection into future prompts
• 3 — Evaluator exists but critique is not looped back to the generator
• 2 — Manual feedback only
• 1 — No feedback mechanism of any kind

6. Error Recovery

Does the system recover from failures without human intervention?

• 5 — Retry with exponential backoff, fallback paths, graceful degradation, error classification
• 4 — Retry logic, most errors handled gracefully
• 3 — Basic try/catch, some retries, errors surface to user
• 2 — Errors surface to user with no recovery path
• 1 — Crashes on any error, requires manual restart

7. Session Continuity

Can the system resume from exactly where it failed?

• 5 — Phase-level checkpoints, auto-resume from exact failure point, state persisted to disk
• 4 — Session state saved to disk, manual resume possible
• 3 — Session state in memory, lost on process crash
• 2 — Must restart from beginning on any failure
• 1 — No session concept, stateless operation

8. Quality Gating

Is there a hard gate that blocks substandard output from shipping?

• 5 — Automated evaluator gate, minimum score threshold enforced, blocking retry with critique injected
• 4 — Gate exists, enforced most of the time, some bypass paths
• 3 — Gate exists but is advisory only, human decides
• 2 — Human review gate only, no automated scoring
• 1 — No quality gate. Anything ships.

9. Cost Efficiency

Is the system cost-optimised for its usage pattern?

• 5 — Zero marginal cost architecture, budget tracking visible per session, cost-per-feature tracked
• 4 — Mostly efficient, minor waste, no tracking
• 3 — Moderate cost, no optimisation attempted
• 2 — Expensive per operation, no cost tracking
• 1 — No cost awareness, potential runaway spend

10. Trigger Automation

Does work start without human initiation?

• 5 — Multiple trigger types all wired: webhook (GitHub/Linear) + cron + Linear poller + Telegram
• 4 — 2–3 trigger types working reliably
• 3 — 1 automated trigger (cron only or webhook only)
• 2 — Manual only, with automation planned but not implemented
• 1 — Fully manual. Human initiates every single run.

11. Knowledge Retention

Does the system get smarter from every run?

• 5 — Lessons stored, analysed weekly, injected into generator prompts for future sessions
• 4 — Lessons stored and sometimes injected into prompts
• 3 — Lessons stored but never surfaced back to the agent
• 2 — Outputs stored but not analysed
• 1 — Nothing is retained between sessions

12. Workflow Standardization

Is the end-to-end workflow defined, documented, and enforced?

• 5 — PITER/ADW templates enforced, intent classifier active, all phases defined, zero ad-hoc paths
• 4 — Workflow defined and mostly followed, minor ad-hoc exceptions
• 3 — Informal workflow, some consistency, often improvised
• 2 — Ad-hoc, workflow varies session to session
• 1 — No workflow standard. Every run is different.

Using This Audit

1. Score each dimension honestly based on what exists in the code, not what is planned
2. Sort ascending — the 3 lowest scores are your highest-ROI targets
3. For each low-scoring dimension, trace back to the specific file or missing file responsible
4. The ZTE level label describes the current state of the system, not its potential