LLM

DKN

Python Agents Just Got Smarter: Agent Skills Is Now Officially Production-Ready

Microsoft has officially released Agent Skills for Python as a stable, production-ready API inside the Microsoft Agent Framework. If you’ve been holding off on building agent-powered applications because of experimental APIs, it’s time to take a second look.

What Are Agent Skills?

Agent Skills is an open format for bundling domain expertise — instructions, reference documents, and executable scripts — into reusable packages that agents load on demand. Instead of cramming everything into a single system prompt, an agent advertises available skill names, then progressively loads only what it needs for the current task. The result: a leaner context window and agents that stay fast and focused.

Each skill is described by a SKILL.md file (for file-based skills) or equivalent code properties. The agent moves through four stages: advertise → load instructions → read resources → run scripts — fetching only what’s relevant for the job at hand.

Three Ways to Build Skills

The release supports three authoring styles, all treated identically at runtime:

  • File-based skills — A directory containing a SKILL.md, optional scripts, and supporting documents. Ideal for cross-functional teams maintaining skills in a shared repository.
  • Class-based skills — Python classes that package instructions and scripts, distributable via internal PyPI feeds like any other Python package.
  • Code-defined skills — Skills created directly in application code, useful when a skill must be generated dynamically or needs to close over application state.

Built for Enterprise Use

Production readiness means more than a stable API. This release ships with the governance controls enterprises need:

  • Human-in-the-loop approval — The three core skill tools require explicit approval by default. Selectively relax approval for trusted, read-only operations.
  • Controlled script execution — File-based scripts are delegated to a runner you supply, giving full control over sandboxing, resource limits, and audit logging.
  • Filtering — Expose only a curated subset of a shared skill library to a specific agent, with context-aware predicates based on the requesting agent or tenant.
  • Caching — Skills resolve once and are reused, with optional per-key isolation for multi-tenant scenarios.

Real-World Use Cases

  • Policy enforcement — Package HR policies, expense rules, or IT security guidelines as skills. Agents load the right policy at query time for consistent, grounded answers.
  • Support playbooks — Turn troubleshooting guides into skills so agents follow documented resolution steps every time.
  • Multi-team composition — Teams author and publish skills independently; you assemble them into a single agent with no cross-team coordination required.

Getting Started

from agent_framework import Agent, SkillsProvider
from pathlib import Path

skills_provider = SkillsProvider.from_paths(
    skill_paths=str(Path(__file__).parent / "skills"),
    disable_load_skill_approval=True,
    disable_read_skill_resource_approval=True,
)

async with Agent(client=client, instructions="You are a helpful assistant.",
                 context_providers=[skills_provider]) as agent:
    response = await agent.run("Help me with onboarding.")

Learn More

With the Python API now stable, teams can build on Agent Skills in production without worrying about breaking changes — a solid foundation for shipping governed, composable AI agents at scale.

Source: Microsoft Agent Framework Blog

DKN

Your AI Coding Agent Can Now Build and Manage Dataverse — Here’s How

The way enterprise software gets built is changing fast. Rather than manually piecing together APIs, command-line tools, and custom scripts, developers are increasingly turning to AI agents — describing what they need and letting the agent figure out the execution. But for that shift to work with platforms like Microsoft Dataverse, the platform itself needs to be something agents can actually understand and operate.

That’s exactly what Dataverse Skills delivers. Released as an open-source plugin for GitHub Copilot and Claude Code, Dataverse Skills gives AI coding agents deep, practical knowledge of Dataverse — from connecting and authenticating to building schemas, loading data, and running analytical queries. All of it driven by natural language.

What Are Dataverse Skills, Really?

At its core, Dataverse Skills is a plugin that teaches your coding agent how to work with Dataverse. It doesn’t expose a menu of commands for you to choose from. Instead, you describe your goal in plain English, and the agent decides which skills to apply, in what sequence, using which underlying tools.

Think of it as giving your AI agent a domain expert’s knowledge of Dataverse — without you needing to be that expert yourself.

The underlying engine uses the Power Platform CLI (PAC CLI) for authentication, solution management, and automation tasks, paired with the Dataverse Web API and Python SDK for data operations. But you never have to think about any of that. Natural language is the only interface you need.

Three Phases, One Unified Experience

The plugin’s capabilities are organized around three core phases of any Dataverse project:

1. Connect

The agent discovers your Dataverse environments, authenticates using PAC CLI or Azure CLI, registers the Dataverse MCP server, and sets up a consistent project structure. You don’t configure anything manually — the agent handles the entire discovery and initialization process.

2. Build

Once connected, the agent can scaffold full data models from scratch: tables, columns, choice fields, lookup relationships, many-to-many relationships, forms, and views. It picks the right tool for each task — MCP for quick reads, the Python SDK for bulk operations, and the Web API as needed — and registers every component into your solution automatically.

3. Operate

With the schema in place, the agent can load data, run cross-table analytical queries, and bulk-import records from CSV files. Need 50 realistic sample records with domain-specific content generated on the fly? One prompt is all it takes.

Seeing It in Action

Here’s a real example of what this looks like in practice. You open your terminal, install the plugin with a single command, and type:

“I’m building a logistics and inventory management system for Veloce Apparel. I need tables for Warehouses, Products, Suppliers, Shipments, and Incidents — with lookups, a many-to-many between Products and Suppliers, and a self-referential shipment routing chain (tracking a package’s journey through hub transfers). Create everything in a VeloceLogistics solution, load sample data, and show me which shipments are currently delayed or stuck in transit.”

From that single prompt, the agent autonomously:

  1. Discovers your Dataverse environment and configures MCP
  2. Creates the solution using PAC CLI
  3. Builds five tables with choice columns, lookups, and a many-to-many relationship
  4. Generates and runs a Python script to bulk-load realistic sample data
  5. Queries across tables to answer the business question

No toggling between documentation tabs. No manual CLI commands. No context switching. The agent orchestrates everything from end to end.

Works With Both GitHub Copilot and Claude Code

Development teams rarely standardize on a single AI coding tool. Some developers prefer GitHub Copilot; others work with Claude Code. Dataverse Skills was built with this reality in mind. Since the skills are written as plain Markdown files with YAML frontmatter — not compiled binaries or proprietary formats — the same plugin works identically in both environments.

Install it from the plugin marketplace for either agent and you get the same knowledge, the same safety checks, and the same results. One investment, both tools covered.

Open Source and Built to Extend

The project is MIT-licensed and openly available on GitHub. Each skill is a standalone Markdown file — readable, editable, and extensible without any compiled code. Teams can add new skills for their own Dataverse customizations, improve existing ones, or contribute bug fixes back via pull request. The architecture is deliberately approachable for anyone who wants to tailor it to their environment.

A Broader Shift in How Platforms Get Used

Dataverse Skills is more than a productivity tool — it signals a broader direction for enterprise platforms. As AI agents become a standard part of the developer workflow, the platforms they interact with need to be operable through intent, not just through traditional interfaces. Describing what you want and having it built, configured, and queryable in your environment is no longer a future concept. With Dataverse Skills, it’s available today.

Getting Started

Install the plugin with one command:

  • GitHub Copilot (VS Code): /plugin install dataverse@awesome-copilot
  • Claude Code: /plugin install dataverse@claude-plugins-official

Then describe your intent and let the agent do the rest.

Original article: Dataverse Skills: Your Coding Agent Now Speaks Dataverse by Suyash Kshirsagar, Microsoft.

DKN

Say Goodbye to Repetitive Admin Tasks: Dataverse Admin Skills Now in Public Preview

If you’ve ever spent an afternoon clicking through the Power Platform Admin Center to apply the same setting across a dozen Dataverse environments, you’ll understand the frustration. Microsoft has now addressed exactly that pain point with the public preview launch of Dataverse Admin Skills — a capability that brings natural language administration to your coding tool, whether that’s GitHub Copilot or Claude Code.

The Admin Bottleneck Problem

Picture this: your security team requests that auditing be enabled across all 20 of your Dataverse environments — today. Without automation, that means 20 separate logins, 20 sets of clicks, and 20 opportunities for human error. Alternatively, you put in a request to a developer to write a bulk script — and wait.

This is the exact gap Dataverse Admin Skills is designed to close. Instead of navigating admin consoles or waiting on scripts, you simply describe what you want in plain English, and the agent handles the rest.

How It Works

Dataverse Admin Skills operates through two complementary paths:

  • Natural Language (Agentic) Path: Using the Dataverse Skills Plugin inside GitHub Copilot or Claude Code, you describe your intent in plain English. The plugin translates your request into the appropriate PAC CLI commands, executes them against the Dataverse Web API, and gives you a clear summary of what changed. It supports multi-environment parallel execution and enforces built-in safety guardrails — including confirmation prompts before any destructive actions.
  • Direct Scripting Path: The same PAC CLI commands powering the agentic experience are available for use in Bash, PowerShell, or SDK scripts. This makes it ideal for CI/CD pipelines, runbooks, and repeatable automation workflows.

Both paths rely on PAC CLI (v2.6+, .NET Framework) and the Dataverse Web API, ensuring a consistent and trusted execution layer.

A Real Example

Say you type: “Enable AllowMCP setting on all environments starting with Preprod.”

Here’s what happens behind the scenes:

  1. The agent lists your Dataverse environments.
  2. It filters for environments matching your criteria.
  3. It asks you to confirm the target list before making any changes.
  4. It updates each environment in parallel.
  5. It presents a summary table of every change made.

One sentence. No browser tabs opened. No scripts written.

What’s Available Right Now

  • Settings Management: Read and update 37 allowlisted PPAC toggles across environments — covering MCP, audit, retention, recycle bin, search, Microsoft Fabric integration, security, and more. Works on a single environment or in bulk with parallel execution.
  • Bulk Delete: Schedule, monitor, pause, resume, and cancel bulk delete jobs. Safety is built in — confirmation prompts, FetchXML validation, and warnings for system tables help prevent accidental data loss.
  • Long-Term Retention: Enable retention on entities, set archival criteria using FetchXML, and track retention jobs. Particularly valuable for compliance scenarios where data needs to be retained but not kept in active storage.
  • Capacity Management (Coming Soon): Storage breakdowns, growth trends, capacity alerts, and archival recommendations — all accessible from your coding tool.

Getting Started in Three Steps

Step 1: Install the Plugin

  • In GitHub Copilot (VS Code): /plugin install dataverse@awesome-copilot
  • In Claude Code: /plugin install dataverse@claude-plugins-official

Step 2: Connect Your Environments

Open your coding tool and ask: “List all my Dataverse environments.” The agent will install PAC CLI if needed, authenticate you, and return your environment list. If anything is missing, it walks you through setup.

Step 3: Try It Out

Here are some prompts to get started:

  • “Enable the Microsoft Fabric integration on all production environments.”
  • “What is the recycle bin retention period for my sandbox environment?”
  • “Disable Dataverse search across all environments in the Europe region.”
  • “Cancel all system jobs that have been stuck in a waiting state since yesterday.”
  • “Set the long-term retention criteria for the custom log table to archive records older than 2 years.”

Why This Matters

Dataverse Admin Skills represents a meaningful shift in how platform administrators interact with their environments. Rather than being constrained by what a UI exposes or waiting for a developer to write automation scripts, admins can now express intent directly — and act on it at scale. The safety guardrails (allowlists, confirmation prompts, parallel execution controls) mean this power comes without sacrificing governance.

This is currently a public preview release, with Microsoft actively refining and expanding the skill set. Now is a great time to explore what it can do.

Original article: Agentic Administration: Dataverse Admin Skills now available in Public Preview by Anirudha Bakore, Microsoft.

DKN

How the Agent-to-Agent (A2A) Protocol is Reshaping Multi-Agent Collaboration

A year ago, Google introduced the Agent-to-Agent (A2A) protocol — a communication standard designed from the ground up for the era of generative AI. Where traditional APIs are rigid and deterministic, A2A was built for agents: fluid, autonomous systems that need to collaborate, hand off tasks, and maintain secure boundaries without getting in each other’s way. As A2A celebrates its first anniversary, the ecosystem has grown far beyond what most anticipated.

The Problem With Treating Agents Like APIs

If you’ve built with AI agents before, you’ve likely hit the wall of trying to wire them together using conventional REST APIs. It works — until it doesn’t. Standard APIs return data or errors. They can’t ask clarifying questions, refine an ambiguous request, or adapt mid-task. And when you start chaining multiple agents together, context windows overflow, proprietary logic leaks, and the whole system becomes a fragile monolith.

A2A was designed to solve these problems at the architectural level. Here’s how:

1. Secure Boundaries — Protecting Your “Secret Sauce”

Enterprise agents often need to work with sensitive internal data or proprietary business logic that should never be exposed to an external system or a public LLM. A2A enables a clean “black box” handoff: you assign a task to a specialized internal agent, it executes in its own secure environment, and only the high-value output is returned. Your data and how-to logic stay encapsulated and private throughout.

2. Zero Context Pollution

Every LLM has a finite context window. Force a primary agent to manage complex, multi-step dependencies on top of a conversation, and you’ll quickly see hallucinations and degraded output quality. With A2A, specialized peer agents manage their own state and dependencies, handling complexity internally without ever crowding the primary agent’s memory. Each participant keeps its focus.

3. Dynamic Autonomy

An API either returns a result or fails. An A2A peer agent does something fundamentally different: it collaborates. It can interpret intent, ask for clarification when the request is incomplete, push back on ambiguity, and adapt its approach based on intermediate results. This transforms inter-agent communication from a data transfer into a genuine working relationship.

4. Distributed Workloads and Modular Design

Instead of one team building an entire agentic solution end-to-end, A2A enables workload distribution. Different components of a solution can be developed and maintained by separate teams, vendors, or managed agentic services — each a domain expert in their slice. The result is a modular architecture that is easier to build, easier to test, and far easier to evolve over time.

Real-World Spotlight: FoldRun and Protein Structure Prediction at Scale

To understand how A2A works in practice, consider one of the hardest problems in biology: predicting a protein’s 3D structure. It requires petabyte-scale genetic databases, specialized GPU infrastructure, and orchestration across multiple AI models (AlphaFold, OpenFold, Boltz). For a developer, building this from scratch is an enormous undertaking.

FoldRun reimagines this entirely. Rather than a fragile pipeline of glued-together APIs, FoldRun is a self-contained, agentic interface. You add it to Gemini Enterprise, the Gemini CLI, or any A2A-compatible environment, assign a structure prediction task in natural language, and FoldRun takes over — managing long-running autonomous tasks, dynamically choosing between models based on prediction confidence, and delivering results as a specialized peer agent. No custom glue code required.

“Having a solution that allows our scientists to use co-folding models with an agentic interface — one which our organization is embracing through Gemini Enterprise — has made testing and integration with workflows much easier.”

— Richard Hughes, BicycleTx

What Else Is the Ecosystem Building?

The A2A community has expanded well beyond life sciences. Here’s a snapshot of where developers are taking the protocol:

  • Agentic Commerce and Autonomous Payments: AI agents are being used to negotiate deals, verify inventory, and execute B2B transactions on behalf of users — with A2A providing the transactional integrity layer.
  • Enterprise Data and Real-Time Streaming: Specialized A2A agents sit at the edge of live event streams and databases, extracting insights and triggering downstream workflows only when specific, compliance-approved conditions are met — without ever exposing raw data to a central model.
  • Cross-Platform IT and DevOps: Operational silos are dissolving. An HR agent can now hand off provisioning parameters to a DevOps peer agent via A2A, which then autonomously configures software licenses, repository access, and secure environments across disconnected SaaS platforms.
  • Secure Telecom and Regulated Networks: In sectors where data exposure is not an option, A2A is being used to implement quantum-safe, end-to-end Message Layer Security (MLS) for autonomous systems — enabling agent collaboration on sensitive data without any underlying information escaping the secure channel.

How to Get Started

The official A2A SDKs are the fastest on-ramp to the ecosystem:

  • Python and Go: Version 1.0 GA — stable and production-ready
  • Java: Beta, tracking the 1.0 spec
  • .NET: Preview, on track for GA
  • JavaScript / TypeScript: Stable on v0.3; 1.0 work in progress

Whether you are building an agent from scratch or extending an existing system to interoperate with the ecosystem, the path to A2A compliance has never been more straightforward.

The Bigger Picture

The shift A2A represents is more than a new protocol. It’s a move away from AI as a monolith — one massive agent trying to do everything — toward AI as an ecosystem: a network of specialized, collaborating agents, each excellent at its job, each secure in its own domain, each able to hand off and receive work through a common language. One year in, that vision is no longer theoretical. It’s shipping code.

Source: Google for Developers Blog — How A2A is Building a World of Collaborative Agents by Alan Blount, Frank Guan, and Nick Losier

DKN

Building Intelligent Agents with Microsoft Agent Framework: The Harness and Claw Explained

What if you could build a fully functional AI agent — one that plans, searches the web, calls your own custom tools, and remembers context across sessions — without wiring together a dozen different libraries? That’s the promise of Microsoft Agent Framework’s harness model, and this post walks through exactly how it works.

This is Part 1 of the Build Your Own Claw series. The guiding example is a personal finance assistant, and by the end you’ll have an interactive terminal agent that can look up stock prices, pull in live market news, and build a step-by-step investment plan on request.

What Is a “Harness” and What Is a “Claw”?

The terminology is deliberately visual. A claw is an agent loop — the thing that wraps a language model, connects it to tools, and keeps the conversation going. A harness is the scaffolding Microsoft Agent Framework provides around that loop: function invocation, conversation history, planning, web search, and file memory are all bundled in by default. You supply the two things that make your agent unique — its purpose (instructions) and its domain-specific tools — and the framework handles everything else.

Step 1: Connect to a Model

Every agent starts with a chat client — the component that communicates with an underlying language model. The framework treats this as a standard interface (IChatClient in .NET), so you can point it at Microsoft Foundry, Azure OpenAI, OpenAI, Anthropic, Google Gemini, Ollama, or any other supported provider without changing the agent code.

For the finance assistant, the setup reads two environment variables — FOUNDRY_PROJECT_ENDPOINT and FOUNDRY_MODEL — and authenticates using Azure’s DefaultAzureCredential. Run az login locally and it just works; in production, swap in a ManagedIdentityCredential for tighter security.

The key insight here: the harness is provider-agnostic. Switching from Foundry to OpenAI is a one-line change in the client setup — the rest of the agent stays the same.

Step 2: Wrap the Client in the Harness

This is where the framework’s value becomes obvious. A single call — AsHarnessAgent() in .NET or create_harness_agent() in Python — transforms the bare chat client into a full agent. You pass in two things:

  • Instructions — a natural language description of what the agent does and how it behaves. For the finance assistant this includes guidance on always verifying numbers with tools, citing sources, and maintaining a watchlist in a watchlist.md memory file.
  • A custom tool — a plain function (get_stock_price) that the model can invoke when it needs live data. The framework automatically generates the JSON schema from the function’s signature and parameter descriptions.

In return, the harness activates everything else automatically:

  • Web search is added as a hosted tool — ask “Any recent news on NVDA?” and it works out of the box, no extra code required.
  • Planning is enabled via a built-in TodoProvider and AgentModeProvider — so a vague request like “Review my watchlist and recommend some stocks to add” becomes a structured, step-by-step plan.
  • File memory is wired up so the agent can persist information (like your watchlist) across sessions.
  • History persistence is handled per service call.

Nothing about web search or planning required any custom implementation — those capabilities came free the moment the harness was created.

Step 3: Run the Interactive Console

Microsoft Agent Framework ships a ready-made terminal UI — the harness console — designed to be copied and adapted as a starting point for any interface (web app, chat surface, IDE extension, etc.). It outputs in colour by mode: cyan for planning, green for execution, and includes built-in commands like /todos, /mode, and /exit.

A typical interactive session with the finance assistant might look like:

  1. Switch to execute mode: /mode execute
  2. Ask for a price: “What’s the price of MSFT?” — the agent calls get_stock_price
  3. Ask for news: “Any recent news on NVDA?” — the agent uses web search
  4. Build a watchlist: “Add MSFT, NVDA and SPY to my watch list” — saved to watchlist.md
  5. Switch to plan mode: /mode plan
  6. Request analysis: “Review my watchlist and recommend some stocks to add” — the agent plans, asks clarifying questions, then executes

Sessions can be saved to disk with /session-export and restored later with /session-import, preserving conversation history, the watchlist, and all context-provider state.

How Plan Mode Actually Works

Plan mode isn’t magic — it’s structured output. When the agent is in plan mode, the console’s planning observer overrides the response format to request a JSON schema-constrained reply instead of free-form text. The model is forced into one of exactly two shapes:

  • Clarification — the model needs more information. It returns one or more questions, each optionally accompanied by suggested choices that the console renders as selectable options.
  • Approval — the model has a complete plan. It returns a single summary and the console prompts you to approve before anything executes.

This design makes agent behaviour feel deliberate and safe: the agent gathers what it needs, presents a plan, and only switches to execute mode after you sign off. The PlanningResponse schema ships with the console sample in both .NET and Python, so you can extend or reshape it to match your own UX requirements.

Toggling Features On and Off

Everything the harness enables — todos, planning modes, web search, file memory, file access, and tool approval — is on by default and individually toggleable. If your use case doesn’t need planning or web search, you disable them with a single option flag:

  • .NET: DisableTodoProvider, DisableAgentModeProvider, DisableWebSearch, DisableFileMemory, DisableFileAccess, DisableToolApproval
  • Python: disable_todo, disable_mode, disable_memory, disable_web_search

The recommended approach is to start with everything enabled and trim to taste once you understand what your agent actually needs.

You Don’t Have to Use the Full Harness

The harness is a convenience layer, not a mandatory container. All of the underlying pieces — web search (a plain tool), planning modes (a context provider), and todos (another context provider) — are individually accessible. You can cherry-pick exactly what you need and plug them into any agent architecture, even one that doesn’t use the harness at all. In .NET, the mode and todo providers live in the Microsoft.Agents.AI package; in Python, everything ships in the agent-framework package.

Try It Yourself

Both the .NET and Python runnable samples are available on GitHub:

What Comes Next

The finance assistant can now look things up, search the web, and produce structured plans. But it can’t touch your files directly, and there’s nothing yet preventing it from taking a sensitive action without asking. Part 2 of the series addresses both: granting file access, gating risky operations behind explicit approvals, and adding durable memory so the agent remembers your preferences between sessions.

Source: Microsoft Dev Blogs — Meet your agent harness and claw by Wes Steyn, Principal Software Engineer

DKN

LangChain Deep Agents: The Easiest Way to Build Reliable AI Agents

Building AI agents that can handle complex, multi-step tasks has never been easier. LangChain’s Deep Agents is a powerful agent harness that brings together everything you need to build reliable, production-ready LLM-powered agents — all in one package.

What Is Deep Agents?

Deep Agents is a standalone Python library built on top of LangChain’s core building blocks. It uses the LangGraph runtime for durable execution, streaming, human-in-the-loop interactions, and more. Think of it as an “agent harness” — the same core tool-calling loop as other agent frameworks, but with built-in capabilities that make agents reliable for real-world tasks.

Whether you’re building a coding assistant, a data analyst, or a research agent, Deep Agents gives you the infrastructure to do it right from day one.

Quickstart

Getting started is straightforward. Install the library and create your first agent in just a few lines:

# pip install -qU deepagents langchain-google-genai
from deepagents import create_deep_agent
def get_weather(city: str) -> str:
    """Get weather for a given city."""
    return f"It's always sunny in {city}!"
agent = create_deep_agent(
    model="google_genai:gemini-3.5-flash",
    tools=[get_weather],
    system_prompt="You are a helpful assistant",
)
# Run the agent
agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)

Core Capabilities

Deep Agents comes with six major built-in capabilities:

  • Take actions in an environment — Invoke tools, read and write files, and execute code.
  • Connect to your data — Load memories, skills, and domain knowledge at the right moment.
  • Manage growing context — Summarize history and offload large results across long runs.
  • Parallelize tasks — Delegate to general or specialized subagents running in isolated context windows.
  • Stay in the loop — Pause for human approval at critical decision points.
  • Improve over time — Update memory, skills, and prompts based on real usage.

Execution Environment

Tools and MCP Support

Pass custom functions, LangChain tools, or tools from any MCP (Model Context Protocol) server using the tools= parameter. Deep Agents fully support MCP, letting you connect to databases, APIs, file systems, and more through a standard interface.

from deepagents import create_deep_agent
agent = create_deep_agent(
    model="anthropic:claude-sonnet-4-6",
    tools=[search, fetch_page, run_query],
)

Virtual Filesystem

The harness provides a configurable virtual filesystem backed by pluggable backends — in-memory state, local disk, LangGraph store, or custom backends. It supports operations like ls, read_file, write_file, edit_file, glob, and grep.

Filesystem Permissions

Declarative permission rules control which files and directories the agent can read or write. You can restrict agents to specific directories, protect sensitive files like .env, and give subagents narrower access than the parent agent.

Code Execution

Deep Agents supports two modes of code execution:

  • Sandbox backends — Expose a shell execute tool for isolated command execution. Ideal for installing dependencies, running tests, or calling CLIs.
  • Interpreters — Add an eval tool running JavaScript in a scoped QuickJS runtime. Great for lightweight data transformations and programmatic tool calling.

Context Management

Skills

Skills package specialized workflows, domain knowledge, and custom instructions for your agent. They follow the Agent Skills standard and use progressive disclosure — the agent reads skill frontmatter at startup and only loads full skill content when a task needs it, keeping startup context compact.

Memory

Memory gives your agent persistent context across conversations — coding style, preferences, conventions, and project guidelines. Memory uses AGENTS.md files and can be updated based on interactions, so preferences carry forward without restating them each session.

Summarization and Context Offloading

The harness automatically compresses conversation history and large intermediate results, isolates subagent work, and uses long-term storage to carry information across threads — all to support multi-step tasks that exceed a single context window.

Prompt Caching

For Anthropic models, Deep Agents automatically applies prompt caching to static sections of the system prompt — base agent instructions, memory, and skill content. This reduces both latency and cost on long-running agents, with no configuration required.

Delegation: Task Planning and Subagents

Deep Agents includes a built-in write_todos tool for structured task tracking with status states (pending, in_progress, completed), giving agents a lightweight planning layer for long-running work.

The subagent system allows the main agent to spin up ephemeral child agents for isolated or parallel tasks. Each subagent gets fresh context, runs autonomously to completion, and returns a single final report — keeping the parent agent’s context clean and token-efficient.

Human-in-the-Loop

Deep Agents integrates with LangGraph interrupts so you can pause for human approval on sensitive tool calls. Use the interrupt_on parameter to specify which tools require a checkpoint:

agent = create_deep_agent(
    model="anthropic:claude-sonnet-4-6",
    tools=[edit_file, deploy],
    interrupt_on={"edit_file": True},  # Pause before every file edit
)

This gives you a runtime safety layer for destructive operations, expensive API calls, and interactive debugging.

Observability with LangSmith

Deep Agents integrates seamlessly with LangSmith for tracing requests, debugging agent behavior, and evaluating outputs. When you’re ready to move to production, LangSmith provides full deployment and monitoring options.

Getting Started

Deep Agents is the right choice if you want a batteries-included agent framework that handles the hard parts — context management, memory, subagent delegation, and human oversight — so you can focus on building what matters.

Source: LangChain Deep Agents Documentation