hermes-esra
AGENTS.md1 starsChanged 2 months agoApache-2.0View on GitHub
Goal: Make Hermes improvement work systematic, value-aligned, observable, and evidence-driven through the Evolutionary Self-Recursive Architecture (ESRA) meta-layer. Official architecture name: ESRA — Evolutionary Self-Recursive Architecture Specification repository: https://github.com/rrpauls/esra This repository: Concrete Hermes implementation of ESRA After a major architecture or skill change, repeated failure, or an explicit full-cycle request, evaluate whether to run one bounded hermes-evolution-orchestrator review. Routine successful work should not trigger a full cycle. A recommendation from evolution_hook.py does not execute a cycle. The caller must explicitly activate…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,774 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository15 other files it gives its agents.
- antifragility-builderSkill
- crisis-managerSkill
- esra-runtimeSkill
- experimenterSkill
- github-actions-integratorSkill
- hermes-codebase-engineerSkill
- hermes-evolution-orchestratorSkill
- loop-auditorSkill
- mental-model-updaterSkill
- ooda-frameworkSkill
- optimizer-philosopherSkill
- self-improverSkill
- self-observerSkill
- system-dynamics-thinkerSkill
- value-clarifierSkill
The file
We split this AGENTS.md into the rules it gives an agent
5 rules. About 0 are common, 1 is distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 15
After a major architecture or skill change, repeated failure, or an explicit full-cycle request, evaluate whether to run one bounded …
- Very specific: line 27
After ./install.sh, Hermes home looks like:
- Very specific: line 60
If unsure, activate the esra-runtime skill or open ~/.hermes/esra/manifest.json.
- Distinctive: line 115
Every 5–10 recorded significant cycles, or after an anomaly, explicitly run the audit.
- Very specific: line 128
All skills live in: ~/.hermes/skills/esra/
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AGENTS.md — ESRA Integration for Hermes
Goal: Make Hermes improvement work systematic, value-aligned, observable, and evidence-driven through the Evolutionary Self-Recursive Architecture (ESRA) meta-layer.
Official architecture name: ESRA — Evolutionary Self-Recursive Architecture
Specification repository: https://github.com/rrpauls/esra
This repository: Concrete Hermes implementation of ESRA
Core Rule
After a major architecture or skill change, repeated failure, or an explicit
full-cycle request, evaluate whether to run one bounded
hermes-evolution-orchestrator review. Routine successful work should not
trigger a full cycle.
A recommendation from evolution_hook.py does not execute a cycle. The
caller must explicitly activate the orchestrator.
Install layout (where things live)
After ./install.sh, Hermes home looks like:
$HERMES_HOME/ # default: ~/.hermes
├── AGENTS.md # this file
├── skills/esra/ # meta-skills (Hermes skill discovery)
│ ├── esra-runtime/ # documents tool paths for the agent
│ ├── hermes-evolution-orchestrator/
│ └── …
└── esra/ # ESRA runtime package (not a skill)
├── manifest.json # machine-readable paths and protocol version
├── conformance.json # evidence-backed capability declaration
└── tools/ # Python CLIs — use these paths| Kind | Path | Discovered by Hermes as |
| Meta-skills | ~/.hermes/skills/esra/ | Skills / slash commands |
| Runtime tools | ~/.hermes/esra/tools/ | Via esra-runtime skill + this file |
| Manifest | ~/.hermes/esra/manifest.json | Read when paths are unclear |
| Conformance | ~/.hermes/esra/conformance.json | Check supported protocol and capability maturity |
Override roots with $HERMES_HOME and/or $ESRA_HOME if set.
Always run tools by installed absolute path (do not depend on a git clone or cwd):
python ~/.hermes/esra/tools/evolution_hook.py
python ~/.hermes/esra/tools/evolution_hook.py --force-cycle
python ~/.hermes/esra/tools/skill_validator.py --verbose --skills-dir ~/.hermes/skills/esra
python ~/.hermes/esra/tools/evolution_dashboard.pyIf unsure, activate the esra-runtime skill or open ~/.hermes/esra/manifest.json.
How to Run the Orchestrator
Recommended Triggers
After this task, run hermes-evolution-orchestrator
orchestrate evolution
run full ESRA cycle
improve the self-development cycle
perform evolutionary audit of this taskSmart Triggering via evolution_hook.py
The helper at ~/.hermes/esra/tools/evolution_hook.py analyzes supplied task
context and local evolution history and recommends whether the caller should
launch the orchestrator.
Callers that preserve lifecycle state must pass origin, root_task_id, and
cycle_depth. origin=esra, any cycle_depth > 0, or a second recommendation
for the same root_task_id is rejected before all trigger heuristics, including
forced or explicit requests. Post-task results are fail-closed: only explicit
success: true is success, and declared artifact paths must exist, be non-empty,
and have a valid image envelope when applicable.
Capabilities:
- Considers task complexity, new skill creation, and result confidence
- Includes rate limiting
- Analyzes historical patterns
- Produces a recommendation and prompt; it is not a native Hermes event hook
Usage:
python ~/.hermes/esra/tools/evolution_hook.py
python ~/.hermes/esra/tools/evolution_hook.py --force-cycleDecision Structuring
For any uncertainty or important decision, run ooda-framework (optionally with related meta-skills such as self-observer, value-clarifier, or loop-auditor).
Periodic Audit
Every 5–10 recorded significant cycles, or after an anomaly, explicitly run the audit. If cycle history is unavailable, report cadence as unknown:
Run loop-auditor to audit the current evolutionary cycleInstalled ESRA Components
Meta-skills
All skills live in: ~/.hermes/skills/esra/
| Skill | Purpose |
esra-runtime | Where tools live — installed paths and how to run them |
hermes-evolution-orchestrator | Central conductor of the ESRA loop |
ooda-framework | Structures work using Observe → Orient → Decide → Act |
self-observer | Honest monitoring of internal state and patterns |
self-improver | Systematic improvement of skills and processes |
value-clarifier | Value alignment checks (mandatory before experiments) |
experimenter | Design and run safe improvement experiments |
mental-model-updater | Integrate results into long-term models |
antifragility-builder | Grow stronger from stress and uncertainty |
loop-auditor | Meta-audit of the entire evolutionary process |
optimizer-philosopher | Deep trade-off and meaning analysis |
system-dynamics-thinker | Feedback loops, stocks & flows, leverage points |
crisis-manager | High-stakes decision making under uncertainty |
hermes-codebase-engineer | Programming and integration work inside Hermes |
github-actions-integrator | CI/CD workflow creation and GitHub Actions automation |
Supporting tools (~/.hermes/esra/tools/)
| Tool | Purpose |
evolution_hook.py | Smart detector + history analysis for triggering the orchestrator |
esra_logger.py | Structured JSON logging for ESRA cycles |
evolution_dashboard.py | CLI visualization of evolution metrics |
baseline_metrics.py | KPI tracking and historical snapshots |
skill_validator.py | Skill frontmatter, DAG, and promotion validation |
experiment_runner.py | Safe experiment lifecycle (canary, staged, A/B) |
hermes_integration.py | Native Hermes hooks, skill injection, feedback loop |
human_oversight.py | GitHub issues/PRs and evolutionary audit trail helpers |
esra_paths.py | Shared path resolution for tools and install layout |
Installation
From the root of this repository:
chmod +x install.sh
./install.shThis installs:
- Meta-skills →
~/.hermes/skills/esra/(includingesra-runtime) - Tools package →
~/.hermes/esra/tools/ - Manifest →
~/.hermes/esra/manifest.json - Conformance declaration →
~/.hermes/esra/conformance.json - This file →
~/.hermes/AGENTS.md
Respects $HERMES_HOME / $ESRA_HOME when set.
Philosophy
- Hermes provides a powerful engine for self-improvement.
- ESRA provides the steering, brakes, navigation, and audit system.
hermes-evolution-orchestratoris the mechanism that connects them.- Skills are discovered by Hermes; tools are a package under Hermes home, advertised by
esra-runtime.
The goal is not merely to add skills, but to make the evolution process itself recursive and self-improving.
Relationship to the ESRA Specification
This repository implements ESRA for Hermes.
The pure conceptual and technical description of the architecture (principles, 8 levels, Loop Execution Protocol) lives in the separate repository:
→ https://github.com/rrpauls/esra
Version: 1.3
Date: 28 July 2026
Compatible with: Hermes Agent + ESRA (Evolutionary Self-Recursive Architecture)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

