Multi-agent crews

Your first crewChapter 8 of 144 min read

Coordinate three specialized agents into a sequential pipeline: one extracts key terms, one evaluates financial risks, and one audits the final memo against the source agreement.

The review team#

Drawing diagram…

InputAgentResultData flows sequentially between agents; dashed lines indicate file reads.

The pipeline separates extraction, evaluation, and verification:

Contract analyst: Extracts commercial terms line by line, citing exact section numbers.

Risk analyst: Analyzes extracted terms from the buyer's perspective and drafts a risk memo.

Fact-checker: Audits every citation, number, and date against the original contract text.

Build the crew#

Implement the multi-agent pipeline in review_crew.py:

review_crew.py

1# review_crew.py
2
3from crewai import Agent, Crew, Process, Task
4from crewai_tools import FileReadTool
5from dotenv import load_dotenv
6
7load_dotenv()
8
9read_file = FileReadTool()
10
11# agents
12
13reader = Agent(
14    role="Contract analyst",
15    goal="Pull every commercially important term out of a contract, with its section number",
16    backstory=(
17        "You read contracts line by line for a procurement team. You quote "
18        "the contract rather than paraphrase it, and you always note the "
19        "section a term comes from."
20    ),
21    tools=[read_file],
22    verbose=True,
23)
24
25risk_analyst = Agent(
26    role="Commercial risk analyst",
27    goal="Spot the deadlines, costs, and one-sided terms that could hurt {customer}",
28    backstory=(
29        "You have reviewed hundreds of software agreements from the buyer's "
30        "side. You know which clauses cost customers money: automatic "
31        "renewals, price increases, short claim windows, and terms the "
32        "vendor can change on its own."
33    ),
34    verbose=True,
35)
36
37fact_checker = Agent(
38    role="Contract fact-checker",
39    goal="Make sure every claim in a contract memo matches the contract word for word",
40    backstory=(
41        "You are the last reviewer before a memo goes to the finance team. "
42        "You check every number, date, and section reference against the "
43        "source document and fix anything that does not match."
44    ),
45    tools=[read_file],
46    verbose=True,
47)
48
49# tasks
50
51extract_terms = Task(
52    description=(
53        "Read the contract at {contract_path}. List its key terms: parties, "
54        "dates, term and renewal, fees and payment, service levels, data "
55        "handling, termination, liability, and any clause that lets one "
56        "party change the deal on its own."
57    ),
58    expected_output="A bulleted list of key terms. Each bullet quotes the contract and ends with its section number.",
59    agent=reader,
60)
61
62assess_risks = Task(
63    description=(
64        "Using the key terms, write a risk memo for {customer}. Pick the five "
65        "terms most likely to cost them money or catch them out, and say what "
66        "they should do about each one."
67    ),
68    expected_output=(
69        "A Markdown memo of no more than 350 words: a two-sentence summary, then "
70        "a numbered list of five risks. Each risk has a section number, why it "
71        "matters, and a recommended action."
72    ),
73    agent=risk_analyst,
74)
75
76fact_check = Task(
77    description=(
78        "Check every claim in the risk memo against the contract at {contract_path}. "
79        "Fix any number, date, or section reference that does not match the "
80        "contract, and remove anything the contract does not say."
81    ),
82    expected_output="The corrected memo in Markdown, followed by a 'Corrections' section that lists each change you made, or says 'None'.",
83    agent=fact_checker,
84)
85
86crew = Crew(
87    agents=[reader, risk_analyst, fact_checker],
88    tasks=[extract_terms, assess_risks, fact_check],
89    process=Process.sequential,
90    verbose=True,
91)
92
93if __name__ == "__main__":
94    result = crew.kickoff(
95        inputs={
96            "contract_path": "contracts/acme-saas-agreement.md",
97            "customer": "Northwind Traders",
98        }
99    )
100    print(result.raw)

Architecture highlights:

Dynamic variables: Placeholders like {customer} resolve across both agent goals and task descriptions.

Selective tool provisioning: Only reader and fact_checker receive read_file; the risk_analyst operates strictly on intermediate context.

Sequential pipeline: Process.sequential chains tasks in declaration order, passing each output to the subsequent agent.

Run the crew#

Execute the multi-agent pipeline:

Terminal

1(env)$ python review_crew.py

Review the final memo and corrections section:

Output

### 5. Limited Liability Cap (Section 7.1)Why it matters: Acme Analytics' liability is capped at the fees paid in the 12 monthsbefore a claim—meaning if the service fails catastrophically (e.g., data breach, systemcollapse), Northwind Traders is limited to $18,000 in damages, regardless of actuallosses....
Corrections:- None. All claims in the memo match the contract word for word.

The crew correctly flags the $18,000 liability cap. However, notice that the fact-checker reports "Corrections: None" despite potential date and calculation risks. In later chapters, you reinforce validation with deterministic Python tools.