Superpower Daily

Superpower Daily

Your reading space

Sign in
Joonko Founder Pleads Guilty to Securities Fraud Over $27 Milli...OpenAI Says It Will Skip a 2026 IPO Over AI Safety ConcernsBaseten Acquires Blaxel to Add Agent Execution to Its AI Infras...LangChain Open-Sources Slack Ad Agent With Human Approval for C...Positron Raises $875 Million to Fund Its Next AI Inference ChipCohere’s Gomez Challenges Lab-Led AI Safety RulesJoonko Founder Pleads Guilty to Securities Fraud Over $27 Milli...OpenAI Says It Will Skip a 2026 IPO Over AI Safety ConcernsBaseten Acquires Blaxel to Add Agent Execution to Its AI Infras...LangChain Open-Sources Slack Ad Agent With Human Approval for C...Positron Raises $875 Million to Fund Its Next AI Inference ChipCohere’s Gomez Challenges Lab-Led AI Safety Rules
Swipe

LangChain Open-Sources Slack Ad Agent With Human Approval for Changes

The tool links ad-platform activity to pipeline outcomes, but its design puts calculations in code and campaign decisions behind a human approval step.

By 2 min read
LangChain Open-Sources Slack Ad Agent With Human Approval for Changes
LangChain Open-Sources Slack Ad Agent With Human Approval for Changes

Listen to this story

The audio brief

About 1:40
0:001:40
Read transcript
LangChain has open-sourced a Slack-based Paid Media Agent that can analyze advertising results and recommend campaign changes, but it stops before touching a live account. Every proposed keyword, targeting adjustment, piece of ad copy, or new search campaign still needs human approval. The agent connects ad-platform activity—spend, impressions, and clicks—with downstream business results from LangChain’s warehouse, including leads, opportunities, and pipeline. Each Monday, it posts a platform-specific report, then answers follow-up questions in the related Slack thread. The more important design choice is where the system draws the line between software and the model. In an early version, the model read raw campaign and pipeline data, performed calculations, and wrote the report. On a frozen test set, that used about 3.9 million input tokens, took 1,112 seconds—roughly 18 minutes—and cost just over three dollars. LangChain says the redesigned workflow cut the run to 85 seconds and made it about 40 times cheaper. Now, code handles repeatable work such as date windows, account matching, calculations, and safeguards. The model interprets performance, explains likely causes, and suggests next steps. LangChain reports that, in its own deployment, paid media reached 20 percent of marketing pipeline in six months, while cost per qualified lead fell 30 percent as spend rose 60 percent. The constraint is that this is a template, not a turnkey marketing operation. It depends on company-specific knowledge, a 19-page wiki, live tools, and warehouse integrations. The key question is whether that code-first, human-approval pattern transfers beyond LangChain’s own setup.

Story brief

3 key points

LangChain has open-sourced its Slack-based Paid Media Agent, a template for connecting ad-platform metrics with leads, opportunities and pipeline data. The system uses deterministic code for calculations, date windows, account matching and safeguards, leaving the model to interpret results and suggest campaign changes. Humans must approve those changes. LangChain reports a reporting run fell from 1,112 seconds to...

  1. 01

    The agent can recommend keywords, targeting, ad copy and search campaigns, but cannot apply changes without human approval.

  2. 02

    An early report used 3.9 million input tokens, cost just over $3 and took 1,112 seconds before redesign.

  3. 03

    LangChain reports paid media grew from 0% to 20% of marketing pipeline in six months; CPL fell 30% as spend rose 60%.

LangChain says it reduced an early paid-media reporting workflow from 18 minutes to 85 seconds. It has now open-sourced the Slack-based Paid Media Agent behind that work, giving teams a starting point for analyzing ad performance and proposing changes without letting the agent apply them on its own.

From weekly report to Slack follow-up

The agent combines advertising-platform data with lead and pipeline data from LangChain’s warehouse. Every Monday, it posts a summary and branded PDF for each platform; the marketing team can then ask it follow-up questions in the corresponding Slack thread.

It can propose keywords, targeting changes, ad copy and new search campaigns. LangChain routes those campaign changes through human approval, making the agent a recommendation system rather than an independent buyer of ads.

The model interprets; code fixes the numbers

LangChain’s initial workflow asked a model to read raw campaign, keyword, pipeline and landing-page data, perform calculations and write the report. On a frozen test set, the company says one report consumed about 3.9 million input tokens, took 1,112 seconds and cost just over $3.

The revised system assigns repeatable work to deterministic code: calculations, date windows, account matching and safeguards. The model instead evaluates campaign performance, explains likely causes and recommends next steps using the results it receives.

The agent’s data boundary

  • Advertising platforms supply media-activity measures such as spend, impressions and clicks.
  • The warehouse supplies downstream outcomes including leads, opportunities and pipeline.
  • Its workspace includes six skills, a 19-page company-specific wiki, live tools and deterministic code.

An internal result, released as a template

The release includes a pattern, not a ready-made marketing operation. LangChain’s own workspace relies on company-specific campaign knowledge and operating rules. Its reported performance gains are from that internal deployment: paid media rose from 0% to 20% of its marketing pipeline in six months, while cost per qualified lead fell 30% from June to August as monthly spend increased about 60%.

LangChain also reports saving about $5,000 a month by bringing analysis and reporting in-house rather than using an agency. The more transferable lesson is its division of labor: systems of record own the metrics, code enforces fixed rules, and people approve changes that affect a live campaign.

Sources

  1. langchain.comHow We Built LangChain’s Paid Media Agent

Loading discussion...

Related coverage

Continue reading