EVOLVE: how ready is your software to evolve itself?

An open, evidence-backed framework and agent skill that answers one question: how ready is a software product to evolve itself safely?

The question breaks into three loops:

Loop	Can the product...
Request → Release	take a user's request, implement it properly and release it?
Issue → Fix	find the problems users face, fix them and ship the fix?
Opportunity → Expansion	notice adjacent needs it does not serve yet, and propose or launch them?

Each loop gets a Software Autonomy Level (SAL) from L0 to L5, modelled on the levels used for self-driving cars:

Level	Name	What the product can do
L0	Manual	Every change is engineers writing and releasing code.
L1	Configurable	People make the change without bespoke engineering, and can roll it back.
L2	Assisted	The product structures the request, diagnoses the issue or drafts the proposal, and drafts the change. People build and release.
L3	Supervised	The product carries the loop end to end; a person approves each release. Every critical control passes.
L4	Policy-bounded	Low-risk changes ship without a person, under policy, with staged exposure, measurement and automatic rollback.
L5	Self-directing	The product initiates, ships and measures change within policy, backed by operational evidence.

Beside each loop's level, EVOLVE reports how far AI itself carries that loop. Each loop is re-scored with its AI-performable stages (turning requests into specifications, building changes, diagnosing issues, proposing fixes, learning from experience, proposing new capabilities) read only on what AI does there. AI first drives a stage at L2, and never above the loop's own level; AI L0 means AI drives no stage.

The headline is the lower of the first two loops, with Expansion reported beside it, for example SAL 2 · AI-driven 0 · Request → Release L2 (AI L2) · Issue → Fix L2 (AI L0) · Expansion L1 (AI L0). The first number says how far the product can evolve itself; the second, how much of that AI does. The scorer always names what blocks the next level, for the product and for its AI.

Behind the levels is the EVOLVE profile: six capabilities, each scored 0 to 4.

	Capability	Asks
E	Elastic (ARC)	Does the architecture scale, survive failure and recover its data?
V	Velocity (DEL)	Can a change be built, verified and released quickly and safely?
O	Open (MAL)	Can the product be reshaped without a code release?
L	Learn (LRN)	Does it notice what is wrong, work out why, and measure whether changes helped?
V	Vet (GOV)	Is every change reviewed, policy-gated, audited and reversible, including the changes it makes to itself?
E	Expand (EXP)	Does it find and launch adjacent value?

As a supporting view, EVOLVE reports AI Readiness for products with AI features: whether the AI features themselves are safe to run in production. It never changes SAL or the AI-driven levels.

Dimension	Checks
Context	grounding in the product's own data (AIR-03); permission-preserving access (AIR-04)
Quality	evaluation sets (AIR-05); regression evaluation before release (AIR-06)
Governance	AI action traceability (AIR-07), plus the AI-specific readings of agent authority, scope limits and untrusted-input handling
Operations	provider resilience (AIR-01), cost controls (AIR-02), quality monitoring (AIR-08), plus the AI-provider part of failure isolation

Each dimension is the floor of its mean; the headline is the weakest dimension, from L0 Foundational gap to L4 Adaptive and operationally proven. Three gates (permission-preserving access, regression evaluation, traceability) cap it at L2, labelled not production-governed. A descriptive AI Capability Footprint shows what the AI does (operate, build, diagnose and improve, expand) with a level per area but no headline; it never changes AI Readiness or SAL.

This is a public beta. It is designed to produce a falsifiable assessment, not a universal software leaderboard. spec/evolve-v0.4-draft.md explains the design and every decision behind it; spec/ai-readiness-v0.5-draft.md covers AI Readiness, and spec/ai-driven-v0.6.md the AI-driven levels.

Field-test results

EVOLVE was run on 11 open-source products from their public source code; each name links to the project, and the field-test README lists the exact commit read. Full tables, evidence and per-product scorecards are in assessments/2026-09-field-test/.

Provisional, single rater. These readings come from one rater and have not been reviewed by the projects' maintainers or by a second rater. Read them as findings to check, not rankings.

Product	Archetype	SAL	Request → Release	Issue → Fix	Expansion	AI Readiness
n8n	configurable platform	2	L2 (AI L2)	L2 (AI L0)	L1 (AI L0)	L2 Deployed with material control gaps
OpenClaw	agent runtime	2	L2 (AI L2)	L2 (AI L0)	L1 (AI L0)	L1 Experimental
PostHog	focused application	1	L2 (AI L2)	L1 (AI L0)	L1 (AI L0)	L2 Deployed with material control gaps
Hermes Agent	agent runtime	1	L1 (AI L0)	L2 (AI L0)	L1 (AI L0)	L1 Experimental
OpenHands	agent runtime	1	L1 (AI L0)	L1 (AI L0)	L1 (AI L0)	L2 Deployed with material control gaps
Dify	configurable platform	1	L1 (AI L0)	L1 (AI L0)	L1 (AI L0)	L1 Experimental
Discourse	focused application	1	L1 (AI L0)	L1 (AI L0)	L0 (AI L0)	no AI by default; L1 Experimental with opt-in
Directus	configurable platform	1	L1 (AI L0)	L1 (AI L0)	L1 (AI L0)	L0 Foundational gap
Frappe	configurable platform	1	L1 (AI L0)	L1 (AI L0)	L1 (AI L0)	no AI features
OpenCode	agent runtime	0	L0 (AI L0)	L0 (AI L0)	L0 (AI L0)	L1 Experimental
LibreChat	focused application	0	L0 (AI L0)	L0 (AI L0)	L0 (AI L0)	L0 Foundational gap

No product's AI drives both the request and the fix loop, so every AI-driven headline is 0.

What stands out:

• AI builds changes, but does not yet fix the product. In n8n, OpenClaw and PostHog, AI turns requests into plans and drafts the change (Request → Release, AI L2). In no product does AI diagnose issues well enough to drive the fix loop. Hermes and OpenClaw come closest: their AI learns from experience and proposes skill changes, but only explains errors when asked.
• No product reaches SAL 3. Definition rollback fails in all eleven: each rolls back some surfaces (workflows, skills) but not all (credentials, memory, configuration).
• No product blocks a release on an AI regression. Every product with AI fails the regression-evaluation gate, so none can reach AI Readiness L3 yet.
• Expansion is the least developed loop. No product clusters unmet demand into themes or proposes an adjacent capability on its own.

What is included

• 66 anchored criteria in six capabilities, each with a 0–4 ladder, plus eight AI Readiness checks (AIR) and AI-qualified readings of 13 existing criteria.

• An AI-driven level per loop: how far AI itself carries the loop, with what it needs for its next level.

• Three loops, each scored by its weakest stage. A loop's level is the lowest of four parts: the loop's own stages, a shared release spine (build, verify, stage, release, observe, roll back), an architecture foundation and a governance ceiling.

• Seven critical controls:

  • definition rollback;
  • release rollback;
  • safe migrations;
  • tested backup and restore;
  • tenant isolation;
  • security;
  • bounded self-change.

  A failed applicable control caps every loop at L2.

• Twelve scope facts (nine for change, three for AI) that decide which criteria and controls apply, so a local tool is not failed on tenant isolation and a product that only changes through code is not failed on definition rollback.

• Evidence grades, evidence facets (implemented, tested, operated) and a depth cap: architecture and control criteria need tested evidence for a 3 and operated evidence for a 4.

• Scoring of the default configuration, with shipped opt-in settings and higher alternate readings reported separately.

• Enforced archetype exclusions and an interpretation guide for agent runtimes and developer platforms.

• A deterministic Python scorer with tests, and a remediation generator that orders work by prerequisites.

• A reusable agent skill (skills/evolvable-software/), packaged as a plugin for Claude Code and Codex.

• A field test on 11 open-source systems in assessments/2026-09-field-test/, scored from public source code and not yet reviewed by their maintainers.

The framework itself contains no private repository evidence, client examples or employer-specific terminology.

Install

Claude Code (plugin):

claude plugin marketplace add harsh51191/evolvable-software
claude plugin install evolvable-software@evolvable-software

Or from inside a session: /plugin install evolvable-software --marketplace harsh51191/evolvable-software. Update later with claude plugin update evolvable-software@evolvable-software.

Codex (plugin):

codex plugin marketplace add harsh51191/evolvable-software
codex plugin add evolvable-software@evolvable-software

Gemini CLI (extension):

gemini extensions install https://github.com/harsh51191/evolvable-software

Cursor, Copilot, Claude Code, Codex and 40+ other agents (via skills.sh):

npx skills add harsh51191/evolvable-software

Any agent that reads skill folders (copy the skill):

git clone https://github.com/harsh51191/evolvable-software.git
cp -r evolvable-software/skills/evolvable-software ~/.claude/skills/   # or ~/.codex/skills/

Restart or refresh the agent after installation. The skill can then be discovered for requests about self-evolution readiness, software autonomy levels, AI-driven change, governed product change, architecture readiness for autonomous change, or agent-safe action surfaces. It works best where it can read the repository being assessed, such as Claude Code, Codex or Cowork.

What the plugin runs and sends

• Reads only the repositories and documents you point it at. It asks before cloning or opening any new remote source, and never changes the repository it assesses.
• Writes one assessment input file (JSON), at a path you choose outside the assessed repository, and the reports you ask for.
• Runs two bundled Python scripts, init_scores.py and score.py in the skill's scripts/ folder, which use the standard library only. They make no network calls and install nothing.
• Sends nothing to any service. The plugin has no hooks, no MCP servers and no telemetry.

Run the scorer directly

Requires Python 3.8 or later. It uses only the standard library.

Create an input template outside the repository you are assessing:

python3 skills/evolvable-software/scripts/init_scores.py \
  --product "Example Software" \
  --archetype focused-application \
  --source "repository @ immutable-tip" \
  --out ~/evolve/example-2026-10-01.json

Every scope fact starts empty and every criterion starts as todo. The scorer refuses the file until each of these holds:

• each fact is true or false, with evidence;
• each criterion is assessed, marked not_evidenced with a search scope, or excluded as the archetype or a scope fact allows.

Complete the file using the skill's references/rubric.md and references/evidence-plan.md, then run:

python3 skills/evolvable-software/scripts/score.py ~/evolve/example-2026-10-01.json

Generate a plan that starts with what blocks the next level:

python3 skills/evolvable-software/scripts/score.py ~/evolve/example-2026-10-01.json --prescribe --target 3

Evidence states and fields

Every criterion must be one of:

• assessed: an anchored score from 0 to 4 with evidence.
• not_evidenced: applicable, but the sources cannot establish presence or absence.
• not_applicable: an exclusion the archetype allows, or a criterion a false scope fact switches off, with a rationale.

Optional fields:

• alt_score: the next level up, also defensible.
• available_score: the level with shipped opt-in settings on.
• facets: implemented, tested, operated.
• inventory: a level per surface, for criteria that span several surfaces.
• if_applicable: what an excluded criterion would score.
• ai: an AI-qualified reading of the 13 criteria that AI Readiness reuses, scored only on AI-backed behaviour. Required when ai_features is true.

Important limits

• Repository evidence establishes that a mechanism exists in the inspected source. It does not prove usability, adoption, hosted-edition parity, production reliability or business impact. The operated facet makes this limit visible.
• A Software Autonomy Level is a readiness reading, not a production certification.
• Cross-product comparison is defensible only when scope, archetype, scope facts, evidence depth and applicability decisions are comparable.
• Give maintainers an opportunity to correct factual evidence before publishing comparative scores.
• Do not include credentials, customer identifiers, confidential excerpts, internal URLs or proprietary evidence in shared scorecards.

Repository structure

.claude-plugin/marketplace.json       Claude Code marketplace entry
.codex-plugin/                        Codex plugin manifest
.agents/plugins/marketplace.json      Codex marketplace entry
skills/evolvable-software/            The skill, which is also the Claude plugin folder
  .claude-plugin/                     Claude plugin manifest and icon
  SKILL.md                            Agent instructions
  references/rubric.md                Criteria, anchored levels and the scoring model
  references/archetypes.md            Applicability and interpretation guidance
  references/evidence-plan.md         Evidence collection plan
  references/remediation.md           Level 3 and 4 improvement moves
  references/scorecard-template.md    Reporting structure
  scripts/init_scores.py              Input template generator
  scripts/score.py                    Validator, scorer and prescriber
spec/                                 Design and decisions for each version
tests/test_score.py                   Scorer tests (python3 -m unittest discover -s tests)
assessments/                          Field-test results
metadata.yaml                         Package metadata
CHANGELOG.md                          Version history

Name and version

EVOLVE v0.6.0, public beta 6, 2 October 2026. EVOLVE was previously the Malleability and Self-Evolution Readiness (MSR) framework; see CHANGELOG.md for what changed and why. "EVOLVE" and "Software Autonomy Level" are working names until a trademark and prior-use search is complete.

License

MIT. See LICENSE.