INTEGRATED
COGNITION.

THE AGENTS FIELD GUIDE

AI agents can act. Judgment still matters.

An agent can use tools and take several steps toward a goal. Its value depends on the task it is given, the information it can access, the actions it may take, and the person responsible for checking the result.

Start with the fundamentals Choose a guide ↓
Diagram of an AI agent moving from a goal through a model, selected tools and observations, with human approval at an action boundary
A useful agent is defined as much by its boundaries as by its tools.

A PATH THROUGH THE SECTION

Understand. Apply. Control. Evaluate.

Begin with the mechanism, then follow an example through the decisions, permissions, and tests that make it usable in legal practice.

  1. 01 / UNDERSTAND

    What is an AI agent?

    Follow the model, tool, and feedback loop. Separate conversation context and stored memory from model training.

    Understand the system
  2. 02 / APPLY

    Agents in legal workflows

    Walk through bounded research, drafting, and review tasks with visible lawyer checkpoints.

    Trace the work
  3. 03 / CONTROL

    Permissions and oversight

    Decide what an agent may read, draft, modify, or send, and when it must stop for approval.

    Set the boundaries
  4. 04 / EVALUATE

    Evaluate an AI agent

    Run a matter-based pilot. Score source support, missed issues, unauthorized actions, and the full time spent on review.

    Use the scorecard

GO DEEPER

Build a fuller mental model.

The four guides above provide the path. These deeper references explain the system underneath it, how it operates interfaces and returns to work, the ways it can fail, and the decisions involved in adoption.

THE PRACTICAL TEST

Four questions before an agent acts.

  1. 01Read

    Which documents, systems, and matter data can it access?

  2. 02Decide

    Which next steps can it choose within a task?

  3. 03Change

    May it alter files, records, or work product?

  4. 04Send

    Can it communicate or submit anything outside the firm?

THE BUILDING BLOCKS

Skills, tools, and handoffs.

Start with the method an agent follows, then explore the tools it can use and the work it can pass to another agent. These three layers explain different parts of the same system.

01 / THE METHOD / AGENT SKILLS

What are Agent Skills?

Agent Skills are an open folder format for giving a compatible AI agent reusable, task-specific instructions. Each skill has a SKILL.md file: YAML frontmatter with a required name and description, followed by Markdown instructions. A skill may also include references, templates, or scripts.

A firm could create an NDA review skill to state its process for finding clauses and reporting issues. The skill supplies a method; the agent still needs authorized source material and a person to check the work.

EXPLORE A FICTIONAL SKILL

An NDA issue-spotting playbook.

Choose a request to see when this example skill might load and where its authority stops. This simulation runs no agent, script, or document review.

EXAMPLE PACKAGEFICTIONAL
nda-review
  • SKILL.mdrequired
  • references/playbook.mdoptional
  • assets/issue-list.mdoptional
  • scripts/compare.pyoptional
INSIDE SKILL.md
---
name: nda-review
description: >-
  Use for NDA clause review; flag issues
  for lawyer review.
---

Compare the supplied NDA with the approved
playbook. Cite clauses and flag uncertainty.
Illustrative files only. A real skill should be reviewed and tested against the firm's current process.

INTERACTIVE EXAMPLE / NO AGENT CONNECTED

When would a skill help?

Choose a fictional request and trace what a compatible agent could load. The choices are predetermined for teaching; no document is read and no skill or script runs.

  1. 01DiscoverName + description
  2. 02LoadSKILL.md instructions
  3. 03CheckSources + resources
  4. 04ReviewLawyer checkpoint
EXAMPLE REQUESTILLUSTRATION ONLY
“Review this supplied NDA against the firm's approved issue list and flag clauses for a lawyer.”

Relevant skill · stage 1 of 4

First, read the skill's label.

A compatible agent can see the skill's name and description before loading the full instructions. Selection depends on the agent and its configuration.

---
name: nda-review
description: >-
  Use for NDA clause review; flag issues
  for lawyer review.
---
EXAMPLE DECISIONDescription fits this request

The request concerns a supplied NDA, which matches the example skill's description. A compatible agent might choose to load its instructions.

AGENT SKILLSThe method

Instructions and supporting files for how to carry out a task.

MCPThe connection

A way for an AI application to use tools or resources from a connected server.

A2AThe handoff

A way to request work from another independent agent system.

Names can mislead. An A2A Agent Card may list an agent's “skills” as advertised capabilities. That description does not mean the agent has an Agent Skills SKILL.md package. A skill package or Agent Card does not grant permission to access client data or replace lawyer review.

Primary references: Agent Skills overview ↗ · Agent Skills specification ↗ · Skills and MCP ↗ · A2A specification ↗

Next, see how MCP connects an agent to tools and resources →

02 / THE CONNECTION / MCP

What is MCP, and how does it connect an agent to work?

Model Context Protocol (MCP) is an open standard for an AI application to connect to outside data and functions. The application is the host. It uses an MCP client to talk to an MCP server, which exposes capabilities from a connected system. A server can run locally or remotely.

In legal work, possible uses include finding a clause in an approved matter folder, placing a tentative deposition preparation hold, or sending a reviewed client update. What the app can request depends on the server and its access, the application's settings, and the connected system's permissions.

Research methods can become tools, too. See Stanford's Paper2Agent example and its possible value for lawyers working with technical evidence.

  • 01 / CONTEXT

    Resources

    Data the application can read, such as a file or a record.

  • 02 / OPERATIONS

    Tools

    Functions the application can call to retrieve information or take an action.

  • 03 / STARTING POINTS

    Prompts

    Reusable templates that help structure a task.

INTERACTIVE EXAMPLE

Follow one request.

Choose a task, then trace how a possible app could handle it. This is a simulation; no system is connected and no request is sent.

  1. 01AI applicationHost + MCP client
  2. 02App boundaryScope + approval policy
  3. 03MCP serverExposes the capability
  4. 04ReviewModel + person

Find a clause · Step 1 of 4

A task becomes a tool request.

The AI application has discovered a tool offered by a connected MCP server. For this example, it prepares the following request:

“Find the termination clause in the approved Matter 42 contract set.”

TOOLsearch_matter_documents

The boundary matters. MCP describes how systems exchange capabilities and results. A connection does not retrain the model or decide which server to trust, what client data may be shared, or whether an answer or action is appropriate. Next, see how agents hand work to one another. Continue with agent architecture and permissions and oversight.

Protocol references: architecture overview ↗ · server capabilities ↗

MCP IN PRACTICE / RESEARCH EXAMPLE

Paper2Agent: from research to callable tools.

An example of turning a research method into tools an agent can call.

Stanford researchers' Paper2Agent connects published research to AI agents. Its 2026 Nature paper ↗ describes a workflow that reads a paper and associated code, prepares the environment, wraps selected runnable methods as MCP tools, and tests them against reference outputs. A compatible agent can then call those tools. The Paper2MCP guide ↗ details the code and testing workflow. This requires usable code; a PDF alone does not become executable software.

The current project guide ↗ separates Paper2Skill, which packages paper context, from Paper2MCP, which exposes existing repository code as tools. See Agent Skills for the instruction layer.

FROM RESEARCH TO A REVIEWABLE RESULT

  1. Paper + usable code

    The method, its dependencies, and reference material.

  2. Prepared MCP server

    Selected tools wrapped and tested against examples.

  3. Compatible agent calls

    Supported operations with specified inputs.

  4. Lawyer + expert review

    Assess assumptions, outputs, and relevance to the matter.

Illustrative workflow, with lawyer and expert review added for the potential legal uses below.

POTENTIAL LEGAL APPLICATIONS

Scrutinize an expert's method.

With a qualified expert, reproduce a published reference analysis or inspect supported parameter changes. Comparing outputs could help a lawyer ask better questions about assumptions and sensitivity.

Examine technical claims.

Technology or patent diligence could include a bounded demonstration of a runnable scientific or AI method, using authorized data and an expert to interpret what the result supports.

Preserve a reviewable record.

Keep source and code versions, inputs, parameters, and outputs for inspection. A tool call alone does not ensure an audit trail; the team must preserve those records.

What still needs review.

The published study ↗ evaluates scientific and computational workflows. It does not validate these potential legal uses. Code, datasets, model files, or dependencies may be missing. Passing covered tests does not establish a method's validity, fit for the matter, or admissibility. Review the method, generated wrappers, license, and execution environment. Use matter data only with authorized access in an approved environment. A lawyer and qualified expert must assess the assumptions, results, and conclusions.

03 / THE HANDOFF / A2A

How A2A lets agents hand off work.

Agent2Agent (A2A) is an open protocol for communication between independent agent systems. A client application or agent can discover a specialist, send a message, follow a longer-running task, and receive its result without knowing how the specialist works internally.

That could help a law firm route a bounded task to a specialist agent. The protocol carries the handoff; the firm remains responsible for choosing the recipient, controlling what information crosses the boundary, and reviewing the work.

MCP

Use a tool or resource

An AI application reaches a defined capability, such as searching an approved document repository.

A2A

Delegate to another agent

An application or agent asks an independent specialist to work on a task and return a message or tracked result.

EXPLORE AN ILLUSTRATIVE HANDOFF

From matter intake to a conflicts review.

Open a stage to see what crosses the boundary. This is a fictional example; no agent is connected and no client information is used.

  1. 01 / DISCOVERFind the specialistAn intake application identifies a vetted conflicts agent.Explore this stageClose this stage

    Its Agent Card advertises its capabilities, service endpoint, and authentication requirements. The card helps the app understand how to reach the agent; the firm must verify the provider and approve the connection separately.

  2. 02 / REQUESTSet a narrow taskThe intake agent asks for possible matches using fictional matter names.Explore this stageClose this stage

    An A2A message can receive an immediate reply or start a tracked task. The firm still decides which names may be shared, who may access them, and whether the receiving agent is authorized for this work.

  3. 03 / RETURNReceive candidatesThe conflicts agent returns possible matches and record references.Explore this stageClose this stage

    For a tracked task, the caller can receive status updates and a result artifact. A match list is evidence to investigate, not a conclusion that a conflict exists or has been cleared.

  4. 04 / REVIEWMake the legal callA lawyer checks the records and follows the firm's conflicts process.Explore this stageClose this stage

    The lawyer examines the underlying records, resolves ambiguous names, and decides the next step under the firm's professional obligations. The protocol does not make that judgment.

TRUST BOUNDARY An Agent Card describes an agent and its connection requirements; it is not proof that the agent is trustworthy. A2A does not grant access to matter data, enforce the firm's approval policy, or decide whether a potential conflict is resolved.

A NAME WITH TWO MEANINGS

Where does ACP fit?

IBM / BeeAI
Agent Communication Protocol

This agent-to-agent project joined A2A in 2025. Its team wound down separate ACP development and moved its work into A2A. It should not be treated as a parallel, current handoff standard.

CODING TOOLS
Agent Client Protocol

A different, active ACP connects code editors and IDEs to coding agents. It serves a developer interface, rather than the independent-agent handoff illustrated above.

Protocol references: A2A core concepts ↗ · A2A specification ↗ · IBM ACP overview ↗ · BeeAI ACP transition ↗ · Agent Client Protocol ↗. Continue to computer and browser use ↓

INTERFACES IN PRACTICE

How agents use computers and browsers.

Computer and browser use let an AI agent operate an interface through tools: observe the current state, choose an action, and check what changed. Browser tools can use page elements or screenshots. Desktop tools can use screenshots, mouse movements, and keyboard input across applications.

A connected service tool requests a defined operation directly. Browser and computer tools work through an interface. One agent may use all three during a task; each needs its own access limits and outcome checks.

Three tool interfaces for retrieving a fictional agreement: a structured service operation, browser page controls, and screenshot-based desktop controls, followed by verification of the saved file.
Conceptual comparison using a fictional task. Browser and desktop controls are tools too; the interface does not determine the agent's authority.

FICTIONAL TASK / NO SYSTEM CONNECTED

One task. Three ways to reach it.

“Find the signed agreement in fictional Matter 42 and save a copy for internal review.”

  1. 01 / SERVICE

    Connected tool

    Request a defined operation with structured inputs and results, through an API or MCP connection.

    IN THIS EXAMPLE

    Search approved Matter 42 records, retrieve the signed version, and return its source reference.

  2. 02 / WEBPAGE

    Browser use

    Read page text and controls, navigate, fill fields, and select elements. Some tools also use screenshots.

    IN THIS EXAMPLE

    Search the approved portal, inspect the agreement row, and select its download control.

  3. 03 / DESKTOP

    Computer use

    Observe screenshots and use mouse and keyboard controls in a permitted desktop environment.

    IN THIS EXAMPLE

    Open the document application, locate the signed agreement, and save an authorized copy.

VERIFY THE OUTCOME

The click is a step. The file is the result.

Check the saved file's matter, version, and signature pages against the request and source. If a download fails or an action's result is uncertain, inspect the current state before retrying. Stop for review when the task is complete or needs broader authority.

Limit the available applications, records, and actions before work begins. Keep unrelated matters and credentials outside the environment. Treat instructions found in pages or documents as untrusted content, and require approval before an external send or other consequential action.

Read the Computer and Browser Use guide

Primary references: Browser-use documentation · Computer-use documentation

WORK THAT RETURNS

Background and scheduled agents.

Background agents carry out a bounded task without keeping a person in an active chat. Scheduled agents start or resume work at specified times; event-driven agents respond to a defined change. These can overlap: a scheduled task may run in the background and use saved progress.

The trigger starts the run. The agent's tools, evidence, and permissions determine what happens next. Between runs, the system can wait without making model calls; a recurring schedule does not give it continuing authority to take new actions.

Schedule, event, and one-off triggers lead to a bounded background run, saved evidence and results, and a next check to report, pause, stop, or resume.
A trigger determines when work begins. Background execution describes how the run proceeds. Saved state helps connect runs; the model need not be active while the system waits.

Three ways to begin the work.

Choose the trigger that fits the task. Each of these can launch background work; a timer or event alone can also launch ordinary automation without an AI agent.

  1. ONCE / BY REQUEST

    Continue this task

    A person starts a bounded job and returns to its result later. It needs a visible status, a way to cancel, and a clear definition of completion.

    FICTIONAL MATTER-FOLDER EXAMPLE

    Prepare a source-linked document index for the approved Matter 42 folder. Stop when the draft is ready for review.

  2. ON A SCHEDULE

    Return at a set time

    A clock starts a new check. Specify the time zone, missed-run behavior, and what happens if the previous run is still working.

    FICTIONAL MATTER-FOLDER EXAMPLE

    Each weekday morning, compare the folder with the last successful check. Record the check time and any new files.

  3. AFTER A CHANGE

    Respond to an event

    A defined event starts work. Confirm the event is authentic and relevant, and handle duplicate or delayed delivery before acting.

    FICTIONAL MATTER-FOLDER EXAMPLE

    When a new document arrives in the approved folder, prepare a draft description with a link to that file.

THE NOTIFICATION RULE

A check should earn your attention.

Give recurring work an owner and an end.

Set the source scope, permitted actions, run and cost limits, notification rule, and stop date before enabling the job. Review access again when the matter or task changes. In this fictional example, the agent drafts internal work for review; it does not send client messages or submit documents.

Read the background and scheduled agents guide

Primary references: LangSmith scheduling · Temporal schedules

KEEP EXPLORING

Examples and perspectives, with their limits in view.

Connect the core guides to a bounded prototype, the assistant concept, and a dated prediction reassessed with current evidence.

PROTOTYPE CASE STUDY

An agentic lawyer demonstration

A video prototype using fictional matters, with its limits and human review points explained.

Read more RELATED GUIDE

AI assistants for professional work

The adjacent concept: assistive tools, multi-step tasks, and why labels alone do not define autonomy.

Read more UPDATED PERSPECTIVE

Can agents do junior lawyer work?

A January 2026 prediction revisited with task-level limits, current examples, and a testing standard.

Read more AGENT ECONOMICS

AI Agent Costs: Virtual Computers, Reasoning, and Scale

AI agent costs split between virtual computers and model inference. See how longer reasoning and parallel agents can grow demand without more users.

Read more