Y

An interactive execute agent with a simple planning loop, tool registry, built-in tools (shell, filesystem, http, python eval), and pluggable model providers (echo, OpenAI, Ollama). Supports streaming and an optional web

Table of Contents

• Overview
• Quick Start
• Agentic Architecture
• Model Routing
• Project Layout
• Development
• Related Repositories

Overview

An interactive execute agent with a simple planning loop, tool registry, built-in tools (shell, filesystem, http, python eval), and pluggable model providers (echo, OpenAI, Ollama). Supports streaming and an optional web

Stack	python
Frameworks	anthropic, fastapi, openai
Tests	yes
Commits	2
Last activity	2026-08-10
Visibility	public

Quick Start

Install

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt   # or: pip install -e .

Run

# Entry point not auto-detected; inspect the layout below.

Agentic Architecture

This repository participates in the sahiixx agentic harness — a shared contract for how agents plan, act, verify, and recover across all repos in this account.

Signal strength: agentic density score 1702 (references to agent, tool-call, LLM, RAG and orchestration primitives across the source tree).

Patterns in play

Pattern	Role here
Prompt Chaining	Deterministic multi-step pipelines where subtasks are known upfront
Routing	Classify input, dispatch to the specialist path (cheap model for easy work)
Parallelization	Independent subtasks fan out; results aggregated programmatically
Orchestrator–Workers	Central planner decomposes dynamically when subtasks can't be predicted
Evaluator–Optimizer	Generator/judge split with explicit rubric; bounded retry
ReAct	Interleaved reason → act → observe for adaptive tool use
Reflection	Self-critique before emitting a final answer

Escalation rule: start with the simplest pattern that solves the problem. Add Reflection only when verification fails, Planning only when dependencies emerge, Multi-Agent only when work exceeds a single role or context window.

Reliability envelope

• Bounded execution — every loop has a max-iteration and wall-clock ceiling.
• Tool sandboxing — filesystem/network side effects are isolated and reversible.
• Guardrail layering — validate at input, mid-loop, and output.
• Context engineering — select, compress, isolate; never let raw history grow unbounded.
• Self-verification — check intermediate output against constraints before continuing.

Model Routing

Agent work in this repo routes through Azure AI Foundry. See AGENTS.md for the full contract.

Purpose	Deployment	Endpoint
Default / general	gpt-5.6-sol	/openai/v1/chat/completions
Deep reasoning	claude-opus-5	/openai/v1/responses only
Embeddings	text-embedding-3-small	/openai/v1/embeddings

export AZURE_FOUNDRY_API_KEY=...        # never commit this
export AZURE_FOUNDRY_BASE_URL=https://<resource>.openai.azure.com/openai/v1

Gotcha: Claude deployments on Azure return 404 api_not_supported on /chat/completions. They answer only via the Responses API.

Project Layout

AGENTS.md
Amp/
Anthropic/
Augment Code/
BEST_PRACTICES.md
CHANGELOG.md
CODE_OF_CONDUCT.md
COMPARISON.md
COMPLETION_REPORT_OCT_2_2025.md
CONTRIBUTING.md
Claude Code/
Cluely/
CodeBuddy Prompts/
Comet Assistant/

Development

# lint / format before committing
ruff check . && ruff format .

# run the CI check locally
gh workflow run hermes-azure-check.yml

Secrets live in environment variables and CI secrets — never in tracked files.

Related Repositories

Part of a 84-repository workspace sharing one agentic contract:

• agentic-harness — patterns, contracts, and reference implementations
• AGENTS.md in every repo pins identical model routing

---

_{README maintained by the agentic harness · last regenerated 2026-08-10}