A Brief, a Gate and a Log: What an Agent-Era Agency Sells, in Code

#ai #agents #python

If agencies sell judgment and a working system instead of hours, the product is a small "pod" setup: a brief an agent can follow, a gate a human must pass, and a log. This post sketches that in Python. It is an illustration, not a customer story, and I have no results to share.

The idea comes from public writing by BCG and Google on agent-native marketing. Both say human marketers move toward direction and oversight, and that work is checked before it goes out.

1. A brief the agent can follow

from dataclasses import dataclass, field

@dataclass
class Brief:
    goal: str
    audience: str
    must_say: list[str]
    must_not_say: list[str]
    owner: str                 # the accountable human
    learning_question: str     # what we want to learn from this run

2. A review gate

@dataclass
class Draft:
    text: str
    brief: Brief
    flags: list[str] = field(default_factory=list)

def check(draft: Draft) -> Draft:
    for phrase in draft.brief.must_not_say:
        if phrase.lower() in draft.text.lower():
            draft.flags.append(f"banned phrase: {phrase}")
    for phrase in draft.brief.must_say:
        if phrase.lower() not in draft.text.lower():
            draft.flags.append(f"missing: {phrase}")
    return draft

3. Human sign-off and a log

import json, time

def release(draft: Draft, approved_by: str, path="release_log.jsonl"):
    if draft.flags:
        raise ValueError(f"fix flags first: {draft.flags}")
    entry = {"ts": time.time(), "owner": draft.brief.owner,
             "approved_by": approved_by, "goal": draft.brief.goal,
             "learning_question": draft.brief.learning_question}
    with open(path, "a") as f:
        f.write(json.dumps(entry) + "\n")
    return entry

What this shows

The agent writes the draft. The code checks it. A named person approves it. The log says who. The four things an agency can still sell sit in these 40 lines: strategy (the brief), system design (the gate), oversight (the approval) and accountability (the owner field).

Limits

This catches only phrase rules. It does not judge tone or truth. A real setup needs a human reader and your own legal review. This is a sketch.

Tej Pandya, founder of GrowEasy.ai.

DEV Community