Abstract

This technical note introduces Earned Agent Autonomy (EAA), a risk-mitigated governance framework for integrating AI agents into production software engineering workflows. EAA posits that operational authority must be treated as a privileged capability, granted incrementally through verifiable performance and objective evidence rather than static model benchmarks. By establishing a structured progression from advisory roles to orchestrated autonomy, organizations can leverage agentic productivity while maintaining systemic integrity and preserving human expertise.

Context & Motivation

The rapid advancement of Large Language Models (LLMs) has led to the deployment of agents with broad system permissions. Current integration patterns often oscillate between restrictive read-only access and permissive "agentic" modes with minimal oversight. This binary approach creates two primary failure modes:

  1. The Senior-as-Janitor Paradox: Senior engineers spend disproportionate time auditing and correcting opaque agent errors, negating the productivity gains of automation.
  2. Systemic Intuition Erosion: Junior engineers risk losing the opportunity to develop foundational debugging and architectural reasoning skills if they supervise systems they cannot yet manually operate.

A more granular, evidence-based authority model is required to scale agentic workflows safely without eroding the human talent pipeline.

Core Thesis

Agent autonomy is a function of trust derived from observed performance, not model capability. Authority should be granted in scoped, revocable increments, gated by the agent's ability to maintain system invariants and operate within explicit execution contracts.

Mechanism: The Autonomy Ladder

Autonomy is managed through five discrete levels. Transitioning between levels requires meeting objective exit criteria based on task history and audit logs.

Level Name Scope Primary Invariants Exit Criteria (Promotion)
L0 Advisory Read-only access to code, documentation, and logs. Information Integrity 0% hallucination rate; 100% citation accuracy over N sessions.
L1 Proposal Generating diffs or PRs for human approval. Architectural Alignment >90% acceptance rate; strict adherence to project style and intent.
L2 Scoped File-system writes within explicit directory bounds. Local Correctness Zero lint/test regressions in scope; predictable rollback behavior.
L3 Task-Bounded Full toolchain access for defined "task contracts." Contract Integrity Successful delivery of N diverse contracts without silent scope expansion.
L4 Orchestrated Goal-driven coordination of multiple sub-agents. Resource & Goal Safety Stable performance under resource bounds; immutable high-level goals.

Concrete Examples

  1. Documentation Maintenance: An agent at L0 identifies outdated README sections based on recent commits. Upon demonstrating 100% accuracy in identifying discrepancies over ten PRs, it is promoted to L1 to propose specific updates.
  2. Test Suite Expansion: An L3 agent is assigned a "Task Contract" to increase coverage for a specific utility module. It operates within a restricted directory, executes local test runners, and is permitted to stage changes only once pre-defined coverage thresholds and performance benchmarks are satisfied.

Trade-offs & Failure Modes

  • Operational Overhead: Implementing EAA requires initial investment in automated audit logging and contract enforcement tooling.
  • False Confidence: Success in narrow, L2-scoped tasks does not guarantee safety in L3-orchestrated tasks involving cross-module side effects.
  • Velocity vs. Rigor: The staged promotion model prioritizes safety over immediate speed, which may conflict with aggressive "AI-first" delivery mandates.
  • Contract Drift: Agents may attempt to bypass constraints by modifying their own execution contracts if those configurations are stored in mutable file paths.

Governance Principles

  1. Enforce Narrow Execution Contracts: Define the minimum necessary file scope and toolset for every agent session.
  2. Log Evidence of Competence: Record the specific successful tasks and invariant checks that justified an agent’s promotion to the next level.
  3. Implement Automated Revocation: Configure "dead-man switches" to automatically downgrade an agent’s level if a critical invariant (e.g., CI failure, security scan hit) is breached.
  4. Guard Human Development: Require junior engineers to supervise L0-L1 agents to ensure they engage in manual reasoning before graduating to L3+ orchestration.

Positioning Note

This model differs from:

  • Academic Safety Research: Focuses on operational implementation rather than theoretical bounds.
  • General Industry Content: Replaces anecdotal productivity tips with a structured governance framework.
  • Vendor Enablement: Remains model-agnostic and emphasizes skepticism and revocability over unconditional enablement.

Status & Scope Disclaimer

This note represents exploratory work within rmax lab. It is a proposed governance model based on internal trials and is not an authoritative industry standard. The effectiveness of this model depends on the specific CI/CD maturity and testing culture of the implementing organization.

Related