hardness1020/learn-agent-architecture
User repository on GitHub
Learn AI agents from scratch.
- Forks
- 109
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Integrations & Adapters
Learn Agent Architecture
A section-by-section guide to agent harnesses - loop, tool runtime, permissions, hooks, subagents, skills, memory, background work - using Claude Code, mini-swe-agent, Hermes Agent, and deepseek-harness as worked examples.
- Stars
- 1.1k
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- hardness1020
- Last verified
- Oct 2026
About this project
Learn how modern AI agents are built around the LLM. English · 繁體中文 · 简体中文 · 日本語 · 한국어. The model reasons. The harness turns that reasoning into controlled action: it runs tools, keeps state across calls, gates side effects, and coordinates loops. A model call cannot do any of those things by itself.
Quoted from the project's README. Read the full README.
- Languages
- Python
Highlights
- Each section follows one lens: the problem it solves, the mechanism, per-system implementations, then failure modes.
- Twenty-plus numbered sections grouped into layers, from the bare agent loop through worktrees, protocols, and autonomy.
- Runnable sections ship source plus a demo script, and the suggested method is to diff a section against the previous one.
- Systems under study are each pinned to a specific version, with more listed as coming later.
Watch out
This is prose plus small runnable demos, not a library: nothing here drops into an application, and the repository-structure and running-the-demos sections fall past the end of the captured README. Each analysis is tied to one specific version of an outside system, so the write-ups drift as those projects ship, and the table itself lists more systems as coming later. The memory layer is explicitly incomplete here, because section 9 continues into the companion learn-agent-memory repository. No benchmarks, interfaces, or adoption guidance are given, and the only licensing statement is an MIT badge. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
The premise is that the model reasons and the harness turns that reasoning into action, and that most agents share one control flow, so the value sits in the layers around it: dispatching tools, gating side effects, managing context, persisting state, coordinating other loops. It argues a coding tool, a chat assistant, and an autonomous runner mostly differ in harness choices, so skim the section list first.
Both, in small doses. The suggested method for a runnable section is to read the loop source and then run its demo, and to diff that section's source against the section before it so the diff is the one mechanism being added. So the code is teaching material sized to isolate a single mechanism, not a framework you would depend on in production.
Not on its own. The section list stops at section 9 for memory, and the README says that section continues in the companion learn-agent-memory repository, which scales the memory loop into a production subsystem across ten more stages. Two further companion repos go deeper on deepseek-harness and on turning real agent cases into reproducible, verifiable eval sets.
Uncertain over time. The systems table pins each study to one version of Claude Code, Hermes Agent, mini-swe-agent, or deepseek-harness, and the harness surface of those tools changes with every release. The README also says more systems can be added later, so treat the section-to-system mapping as a reading guide and re-check any specific detail you intend to rely on.
Each study is pinned: Claude Code at v2.1.88, Hermes Agent at v2026.7.1, mini-swe-agent at v2.4.5, and deepseek-harness at dsh-v0.1.0-rc.7. The table also lists which sections apply to which system, since they are not all studied end to end: Hermes Agent is read for sections 7, 9, 14, 16, 19, 21, and 22, while Claude Code covers all of 0 through 23.
Twenty numbered sections across eight layers. Layer 1 is the core loop: agent loop, tool runtime, permission and sandbox, hooks. Layer 2 adds planning and todos, subagents, skills, and context management. Layer 3 covers memory, system prompt assembly, and error recovery. Layer 4 is task system, background execution, scheduling, and worktree isolation. Layer 5 is coordination, protocols, and autonomy. Layer 6 covers MCP, plugins, and channels.
No. This is general agent-harness architecture with no Jev content, and it is included as reading rather than as a Jev tool. What it gives you is the vocabulary for the layers a decision model would sit inside, which is why it is useful next to the guard, compaction, and routing entries rather than instead of them.
More Jev resources like this
25GitHub stars
Agent runtime with a terminal UI and a messaging gateway that keeps memory, self-built skills, session search, scheduling, and subagents across Telegram, Discord, Slack, WhatsApp, Signal, and CLI.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
wundercorp/lokiPython
wundercorp/lokiPython51GitHub stars
GraphRAG layer that runs a 4-phase ingest-to-answer pipeline over Neo4j, Memgraph, AGE or Kuzu using Laya or Jev decisions instead of per-hop LLM calls.
Jev Integrations & Adapters#community#jevbest#catalog-import
5GitHub stars
Community Agent Zero plugin that exposes one typesafe_query tool for combined Choice, Noul, and Score questions over the same evidence, with probability cards in chat.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
3clyp50/a0-typesafe-aiPython
3clyp50/a0-typesafe-aiPython3GitHub stars
A Home Assistant Assist conversation agent that sends each command to Jev as one request with a speculative fan-out of questions, then hands the typed answers to Home Assistant's own intents to do the device control.
Jev Integrations & Adapters#community#jevbest#catalog-import
Community
JanOstrowka/typesafe-assistPython
JanOstrowka/typesafe-assistPython20.5kGitHub stars
Python agent framework whose TypeSafeModel class runs Jev as a decision model inside Pydantic AI agents: Pydantic output-type fields become typed Jev questions answered together in one request, each answer carrying its own confidence, and tools can be offered as Choice options.
Jev Integrations & Adapters#community#agents#framework
Community
pydantic/pydantic-aiPython
pydantic/pydantic-aiPython691GitHub stars
A Python speech-to-speech avatar framework: swappable VAD, STT, LLM, and TTS modules behind one streaming pipeline, with WebSocket, telephony, LINE, and OpenAI-compatible channel adapters on top.
Jev Integrations & Adapters#community#jevbest
Community
uezo/aiavatarkitPython
uezo/aiavatarkitPythonFrom the Jev community
Posts from builders shipping with Jev right now.