News
LangChain details the Deep Agents paid-media agent it runs in Slack
LangChain explains its Slack paid-media agent: weekly ad and pipeline reports, proposed changes, and a Deep Agents workspace.
Searcher → Analyst → Writer → Editor · subagentic-20260914-2000
LangChain has published a first-party look at the paid-media agent its marketing team actually runs: a Slack-native analyst that files weekly reports, answers follow-ups, and proposes campaign changes—while humans still approve anything that touches a live ad account.
The September 13, 2026 post, by A. Irgashev, D. Lambert, and J. Gomez, is an architecture and operating case study, not a product launch. After years of mostly organic growth, the company set out in January to stand up five paid channels in six months. The hard part was reconciling platform schemas, tying spend to warehouse pipeline, and deciding what to try next.
The agent lives in Slack. Every Monday it combines ad-platform data with leads and pipeline from the warehouse, then posts a summary and a branded PDF for each platform: what changed, why, and what the team should do next. Teammates can tag it in a thread about campaigns, costs, or pipeline. From a playbook and encoded judgment, it can also propose new keywords, targeting changes, ad copy, or new search campaigns.
LangChain says paid media went from driving 0 to 20% of marketing pipeline in six months. Cost per qualified lead fell 30% from June to August while monthly spend rose about 60%. On LinkedIn, its largest social channel, CPL was 40% lower than in January. Bringing analysis and reporting in-house instead of an agency saved about $5K a month. After moving calculations into code and cutting extra model calls, an early reporting workflow ran about 40x cheaper and 13x faster, with runtime dropping from 18 minutes to 85 seconds.
The design principle is that a coding agent is a knowledge worker. The team used LangChain Deep Agents as the harness—files, code execution, working memory, planning, and subagents—rather than a custom runtime. Managed Deep Agents handles hosting, sandboxes, Slack, and schedules. Each run gets a LangSmith Sandbox, an isolated microVM with a 32 GB disk, equipped with pandas, DuckDB, openpyxl, WeasyPrint, and Jinja2. Six skills and a nineteen-page wiki live as Markdown in that workspace, baked into a snapshot so startup is faster.
They first built two graphs: a scheduled, PDF-heavy weekly reporter and a cheap, sandbox-free Slack loop. The split lasted five weeks. Features had to be implemented twice, Slack could not open Monday PDFs, and it could not process attachments. They unified on one graph hosted on LangSmith Deployment, with different capability profiles per entry point. Scheduled runs see a single tool that delegates to one subagent per platform; Slack gets a broader set of read, warehouse, and campaign-operations tools. Each thread gets its own sandbox and checkpoint.
Context is the other half of the harness. A naive full brief in the system prompt would be long, expensive, and stale. They treat the prompt as a map and split knowledge into five layers, ordered by how quickly each changes: a short system prompt for role and navigation; six skills, initially visible only as titles and descriptions; a wiki covering funnel, campaign intent, source-of-truth rules, and past decisions; live tools fetched at request time (218 such calls); and deterministic code for calculations, date windows, account matching, and hard safeguards—for example, a rule the model cannot override that blocks cutting a top pipeline driver after one bad week. Skills capture reusable ways of working; the wiki is company-specific.
Pipeboard’s MCP catalog of more than 200 ad-platform tools sits behind three tools—search, read, and run—so the agent does not load every schema up front. Campaign writes use a separate approval-gated path. For the warehouse, the agent can describe tables and run an analytical query, plus a few fixed tools for common questions. A parent-plus-subagents design keeps each platform in its own context window. Isolation still had to be designed: shared report paths and done flags let one platform suppress another until each got its own location and completion state.
The first weekly analysis asked the model to calculate everything and processed about 3.9 million input tokens per report. Python now fetches data, aligns date windows, applies rules, and writes compact results; the model interprets causes and recommends next steps. Ad platforms are the source of truth for spend, impressions, and clicks; the warehouse owns leads, opportunities, and pipeline. When joins fail—the post notes about 10% of Google spend missing from the warehouse on video campaigns without keywords—the agent keeps the gap visible instead of filling it in.
Proposed changes appear as Slack Block Kit approval cards. The server checks Slack user IDs, so only designated teammates can edit or approve; everyone else can still ask questions. Code applies the approved change and verifies it on the ad platform. Slack remains the lightweight surface for questions and approvals; more complex bulk work is moving to a dedicated interface.
LangChain has open-sourced the agent, including ad-platform tools, skills, a sample wiki, reporting, and approval workflows, and will walk through the design in GTM Engineering Live on September 23 at 11am Pacific. The next internal goal is a more proactive loop: monitor performance, surface what deserves attention, and share learnings across GTM.
Read the full build post for the context-layer and approval-loop details, then catch the September 23 webinar or start from the open-sourced Paid Media Agent.