Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 

README.md

161 — Agent Harness Engineering: LangChain DeepAgents, Virtual Filesystems & the Ralph Loop


topic: agents type: research status: research-complete last-validated: 2026-05-21 original-query: Analyze agent harness engineering framework (LangChain DeepAgents) and extract patterns for ZAO OS agent roadmap (reconstructed) tier: reference

Status: Research complete Date: March 27, 2026 Goal: Analyze Viv Trivedy's (LangChain) agent harness engineering framework and extract patterns applicable to ZAO OS's AI agent roadmap Source tweet: x.com/vtrivedy10/status/2037203679997018362


Key Decisions / Recommendations

Decision Recommendation
Core concept Agent = Model + Harness. The harness (system prompt, tools, filesystems, sandboxes, orchestration) matters as much as the model itself
For ZAO OS ElizaOS agent Apply harness engineering principles — don't just pick a model, design the system around it (tools, memory, context strategy)
Virtual filesystems USE virtual filesystem pattern for any ZAO agent that needs persistent state across sessions (onboarding agent, moderation agent, fractal tracking)
Ralph Loop for long tasks For any ZAO agent doing multi-step work (e.g., weekly fractal compilation, cross-platform publishing), implement a Ralph Loop: intercept completion, relaunch in clean context with filesystem state
Context rot prevention ZAO OS agents will need compaction (summarize old context) + tool output offloading (save verbose results to filesystem, keep only summaries in context)
DeepAgents library EVALUATE langchain-ai/deepagents as the runtime for ZAO's planned AI agents — it provides planning, filesystem, subagents, and context management out of the box
Current ZAO OS state No harness/agent infrastructure exists yet in src/. This is greenfield. Memory doc references ElizaOS agent as post-client priority

The Harness Engineering Framework

Agent = Model + Harness

Viv Trivedy (@vtrivedy10, LangChain) has been publishing a series on harness engineering — the discipline of building structured systems around LLMs to make them reliable work engines. The core insight:

"If you're not the model, you're the harness."

The harness includes everything that isn't the model weights:

  • System prompts and tool descriptions
  • Skills, MCPs, and bundled infrastructure
  • Filesystem and sandbox environments
  • Orchestration logic for subagent coordination
  • Middleware hooks for deterministic execution

Key finding (May 2026): LangChain's coding agent improved from 52.8% to 66.5% on Terminal Bench 2.0 (rank Top 30 → Top 5) through harness engineering alone, without model changes. On Terminal Bench 2.0, Anthropic's Opus 4.6 performs significantly better in optimized third-party harnesses than in native Claude Code. Infrastructure design rivals model selection in importance.

Six Core Harness Components

1. Virtual Filesystems (Durable Storage)

Filesystems are the foundational agent pattern. LangChain uses them for:

  • Data storage — persistent workspace across sessions
  • Memory — agent knowledge that survives context window resets
  • Collaboration scratchpad — shared state between subagents
  • Git integration — versioning, branching, rollback for agent work

Any storage can "look like a filesystem" by mapping read, write, ls to a backend (S3, Notion, Box, Supabase, custom). Agents stay in their training distribution (great at filesystem ops) while plugging into real-world storage.

DeepAgents tools: read_file, write_file, edit_file, ls, glob, grep

ZAO OS relevance: A fractal-tracking agent could use a virtual filesystem backed by Supabase to persist weekly results, member scores, and submission history — all accessible via familiar file operations.

2. Bash and Code Execution

Rather than pre-configuring every possible tool, harnesses grant agents general-purpose code execution. This enables autonomous problem-solving without human intervention for each action type.

3. Sandboxes (Secure Environments)

Isolated execution environments protect against malicious code while enabling:

  • Dependency installation
  • Environment scaling
  • Tool pre-configuration
  • Self-verification loops through testing and log inspection

4. Memory and Continual Learning

Memory files (like AGENTS.md, CLAUDE.md) inject contextual knowledge at startup. Combined with web search and MCP tools for accessing information beyond training data.

ZAO OS relevance: ZAO already uses CLAUDE.md + community.config.ts for context injection. This same pattern extends to any AI agent — inject community config, member data, governance rules at startup.

5. Context Rot Mitigation

As context windows fill, agent performance degrades ("context rot"). Solutions:

Strategy How It Works ZAO OS Application
Compaction Summarize old context before window fills Long governance discussions, fractal history
Tool output offloading Save verbose results to filesystem, keep summaries in context API responses, member data queries
Skills/progressive disclosure Load capabilities on-demand, not all at startup Agent loads music tools only when in music context

6. The Ralph Loop (Long-Horizon Execution)

The most novel pattern. For tasks spanning multiple context windows:

  1. Agent works toward a goal using filesystem for state
  2. When the context window fills or agent tries to exit prematurely
  3. A harness-level hook intercepts the exit
  4. Reinjects the original prompt in a clean context window
  5. Agent picks up from filesystem state and continues

This enables multi-hour, multi-window autonomous work. The filesystem is the bridge between context windows.

ZAO OS relevance: A weekly fractal compilation agent could use the Ralph Loop to:

  • Process 100+ weeks of historical data across multiple context windows
  • Compile cross-platform publishing reports
  • Generate governance summaries from multi-day proposal discussions

Training-Harness Co-evolution

Models post-trained alongside specific harnesses develop specialized competence within those systems. But this can create brittle overfitting. The research shows harness optimization alone (without model retraining) can deliver significant performance gains.


LangChain DeepAgents: The Implementation

langchain-ai/deepagents is LangChain's open-source agent harness library.

Features

Feature Description
Planning write_todos tool for task decomposition and progress tracking
Filesystem read_file, write_file, edit_file, ls, glob, grep
Shell execution execute command with sandboxing
Subagents task tool for delegating work with isolated contexts
Context management Auto-summarization for long conversations
Smart defaults Built-in prompts teach models effective tool usage
Any model Works with any LLM that supports tool calling

Quick Start

from langchain.chat_models import init_chat_model
from deepagents import create_deep_agent

agent = create_deep_agent(
    model=init_chat_model("anthropic:claude-sonnet-4-6"),
    tools=[my_custom_tool],
    system_prompt="You are a ZAO community assistant.",
)

Virtual Filesystem Backends

Storage can be routed to different backends:

  • Local filesystem
  • S3
  • Notion
  • Box
  • SQLite
  • Supabase (relevant for ZAO OS)

LangSmith Fleet (March 19, 2026)

LangChain also launched LangSmith Fleet with primitives for agent-heavy futures:

  • Agent Identity — identity + security models for agents doing work on behalf of humans
  • Observability, tracing, cost tracking
  • Prompt management and versioning

Evaluation: HumanEval Benchmarks

Three harness variants tested on coding problems:

  1. Basic prompt — baseline
  2. Self-verification — "mentally run examples, rewrite if failing"
  3. Step-by-step reasoning — explicit corner-case analysis before coding (4/5 problems solved)

Adding ModelCallLimitMiddleware (max 2 invocations per call) reduced costs while maintaining results. Harness constraints can improve both cost and reliability.


Related Viv Trivedy Posts (March 2026)

Date Topic Key Insight
Mar 10 The Anatomy of an Agent Harness Agent = Model + Harness; harness engineering > prompt engineering
Mar 10 Decoupling Storage from Compute Give each subagent dedicated compute, share storage
Mar 15 DeepAgents library Internal agents at LangChain built on it
Mar 19 LangSmith Fleet launch Agent Identity for agent-heavy workflows
Mar 25 Virtual filesystems Filesystem pattern for data storage, memory, collaboration
Mar 26 X Article (this tweet) Comprehensive harness engineering article (1.2K likes, 182 RTs)

How This Maps to ZAO OS

Current State (March 2026)

  • No AI agent infrastructure in src/
  • ElizaOS agent planned (memory doc: project_elizaos_agent.md) for onboarding, support, context help
  • CLAUDE.md + community.config.ts already implement the "memory file" harness pattern
  • Supabase available as a virtual filesystem backend

Recommended Agent Architecture for ZAO OS

ZAO Agent Harness
├── Model: Claude (via Anthropic API or DeepAgents)
├── Harness:
│   ├── System prompt: community.config.ts + member context
│   ├── Virtual filesystem: Supabase-backed (proposals, fractals, members)
│   ├── Tools: governance queries, music curation, cross-post publishing
│   ├── Memory: AGENTS.md with ZAO rules + rolling conversation summaries
│   ├── Context management: compaction + tool output offloading
│   └── Ralph Loop: for multi-step tasks (fractal compilation, report generation)
└── Subagents:
    ├── Onboarding agent (member intake, wallet verification)
    ├── Moderation agent (Perspective API + community rules)
    ├── Fractal agent (weekly compilation, score tracking)
    └── Publishing agent (cross-platform content distribution)

Key Takeaway for ZAO

Don't just "add an AI agent." Design the harness:

  1. What tools does it need? (Supabase queries, Neynar API, publishing APIs)
  2. What state persists across sessions? (member data, governance state, fractal history → virtual filesystem)
  3. How do you prevent context rot? (compaction for long governance threads)
  4. What happens when a task spans multiple windows? (Ralph Loop for weekly reports)
  5. How do subagents coordinate? (shared Supabase-backed filesystem, isolated compute)

Sources [FULL - May 21, 2026]

Updated 2026-05-21: LangChain's breakthrough results confirmed - 52.8% to 66.5% improvement on Terminal Bench 2.0 (gpt-5.2-codex) through harness engineering alone. Three key improvements: system prompts with self-verification loops, enhanced tools + context injection, middleware hooks for doom-loop detection (PreCompletionChecklistMiddleware). Validates that agent infrastructure design rivals model selection in importance.