1
# Every agent response must be auditable
2
class BankingAgent:
3
    def __init__(self):
4
        self.audit_log = []
5

6
    def run(self, query: str):
7
        # Log the raw query
8
        self.audit_log.append({
9
            "timestamp": datetime.utcnow().isoformat(),
10
            "type": "user_query",
11
            "content": query,
12
        })
13

14
        # Generate response with explicit reasoning
15
        response = self.agent.run(f"""
16
        <task>{query}</task>
17

18
        <compliance_rules>
19
        - Do NOT process transactions over $10,000 without verification
20
        - Do NOT share account details of other users
21
        - Do NOT disclose internal fraud detection logic
22
        - All responses must include a disclaimer for financial advice
23
        - If unsure, escalate to human agent — do not guess
24
        </compliance_rules>
25

26
        <audit_required>
27
        Before responding, explain your reasoning in this audit block
28
        (the customer will NOT see this):
29
        - What information did you use?
30
        - What compliance rules apply?
31
        - Did you verify the customer's identity?
32
        </audit_required>
33

34
        <customer_response>
35
        Your response to the customer goes here.
36
        Include the standard disclaimer.
37
        </customer_response>
38
        """)
39

40
        # Audit the full response
41
        self.audit_log.append({
42
            "timestamp": datetime.utcnow().isoformat(),
43
            "type": "agent_response",
44
            "content": response,
45
            "compliance_check": self.verify_compliance(response),
46
        })
47

48
        return response

1
# The agent doesn't ask for everything upfront
2
verification_levels = {
3
    "balance_check": {"name", "last_4_digits"},
4
    "transaction_history": {"name", "last_4_digits", "dob"},
5
    "transfer": {"full_ssn", "2fa_code", "device_verification"},
6
    "account_changes": {"full_ssn", "2fa_code", "manager_approval"},
7
}
8

9
agent = Agent(
10
    instructions=f"""Verify identity progressively based on the action:
11
    {json.dumps(verification_levels, indent=2)}
12

13
    For balance_check: ask only name + last 4 digits.
14
    For transfers: require SSN + 2FA + device verification.
15
    Never ask for more information than needed.
16
    If verification fails at any level, do not proceed.
17
    """
18
)

1
support_agent = Agent(
2
    instructions="""You are a SaaS support agent.
3

4
    Response structure:
5
    1. **Acknowledge**: Show you understand the problem
6
    2. **Diagnose**: Ask clarifying questions if needed
7
    3. **Resolve**: Provide step-by-step solution
8
    4. **Confirm**: Ask if the solution worked
9

10
    Escalation rules:
11
    - If the user is frustrated (uses caps, swears, repeats), skip diagnosis
12
    - If you can't resolve in 3 messages, offer to escalate
13
    - If the issue affects billing, always loop in billing agent
14
    - Never make the user repeat information they already provided
15

16
    Tone rules:
17
    - Professional but warm. Use "I" not "we".
18
    - Never blame the user: "Let's fix this together" not "You did X wrong"
19
    - If you don't know something, say "I don't have that information" — don't guess
20
    """,
21
    handoffs=[billing_agent, technical_agent, human_agent],
22
)

1
# Onboarding changes as the user progresses
2
onboarding_stages = {
3
    "welcome": {
4
        "instructions": "User just signed up. Be warm. Guide them to create their first project. Offer a quick tour. Don't mention billing yet.",
5
        "max_messages": 3,
6
    },
7
    "active": {
8
        "instructions": "User has created a project. Help them configure features. Ask about their goals. Offer relevant tips based on their actions.",
9
        "max_messages": None,
10
    },
11
    "power_user": {
12
        "instructions": "User has been active for 30+ days. Skip basics. Offer advanced features, API access, and integrations. Ask if they want a demo of new features.",
13
        "max_messages": None,
14
    },
15
    "churning": {
16
        "instructions": "User hasn't logged in for 14+ days. Be proactive but not pushy. Ask what's preventing them from using the product. Offer a call with customer success.",
17
        "max_messages": 2,
18
    },
19
}

1
dev_agent = Agent(
2
    instructions="""You write and modify code.
3

4
    Mandatory workflow:
5
    1. ALWAYS read the relevant file first (use read_source_file tool)
6
    2. ALWAYS check existing tests before adding features
7
    3. ALWAYS run tests after changes
8
    4. NEVER overwrite a file without reading it first
9
    5. NEVER delete code without understanding what it does
10

11
    If you can't read the file (permission error, not found):
12
    - Do not guess the content
13
    - Report the error to the user
14
    - Ask for the correct file path
15

16
    Deployments (CRITICAL):
17
    - Never deploy on Friday after 2 PM local time
18
    - Never deploy without a PR review
19
    - Always run the test suite before deployment
20
    - If tests fail, fix them before deploying
21
    """,
22
    mcp_servers=[github_mcp, filesystem_mcp, testing_mcp],
23
    system_prompt="You are a senior developer. Read before writing. Test before deploying."
24
)

1
pr_review_agent = Agent(
2
    instructions="""Review pull requests for this repository.
3

4
    Checklist (ALL items must pass):
5
    1. [ ] Tests pass (check CI status)
6
    2. [ ] No secrets committed (check for API keys, tokens)
7
    3. [ ] Code follows project conventions (check steering file)
8
    4. [ ] No commented-out code
9
    5. [ ] Error handling is present
10
    6. [ ] No console.log or debug statements in production code
11

12
    Scoring:
13
    - 0-2 failures: Approve with comments
14
    - 3-4 failures: Request changes
15
    - 5-6 failures: Block and tag team lead
16

17
    Tone: Specific, actionable, and respectful.
18
    "Line 42: This variable is unused" not "Your code is messy".
19
    """,
20
    tools=[read_file, search_code, check_ci_status],
21
)

1
// Bad: 2,500 words covering project rules, identity, tools, output format, safety, examples, and the meaning of life

1
// Bad: Critical rule hidden in paragraph 47 of 83
2
"You must never delete production data... (scrolling)... and also please format dates as ISO 8601"

1
# Bad: Agent fails silently
2
agent.run("Deploy to production")  # Agent decides not to deploy but doesn't say why

1
instructions = """If you decide NOT to do something, explain why.
2
Silent refusals are unacceptable — tell the user what's blocking you."""

1
# Bad: Same instructions, different names
2
agent_a = Agent(name="researcher", instructions="Be thorough and accurate")
3
agent_b = Agent(name="writer", instructions="Be thorough and accurate")

1
// Bad: Project rules that change every sprint are in a static system prompt

1
// Bad: Prompt tells the agent to do something a tool could do
2
"Calculate the total by multiplying price by quantity for each item and summing"

1
// Bad: The model complies but the content suffers
2
"Format: strict JSON with exactly these 47 fields..."
3
// Result: Valid JSON, hallucinated data

1
// Bad: Deployed a prompt change and "it felt better"

1
// Bad: "I just tweaked the instructions in production" → 2 hours later: P0 incident

1
// Bad: Tool descriptions + system prompt + steering file consume 60K tokens
2
// The model has 12K left for the actual conversation

1
# Bad: Same prompt for all users, all contexts
2
agent = Agent(instructions="Help the user with anything")

1
// Bad: "The new prompt broke everything but we don't have the old version"