Multi-Step GTM Agent for Sales Lead Processing and Account Intelligence

langchain · Tech · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Tech

LangChain wanted to reduce GTM manual work around lead research, outreach drafting, and account intelligence while keeping sales representatives in control of messages.

Constraints:

long-running multi-step sales workflows
many internal and external data sources
representatives must review outreach before it is sent
quality needs to improve from logged edits and outcomes

Architecture

LangChain's GTM agent is a long-running sales workflow built on Deep Agents, connected to Salesforce, Gong, LinkedIn/web research, Slack, LangSmith, PostgreSQL memory, and compiled subagents. It triggers from Salesforce leads, performs do-not-send checks, gathers context, drafts personalized email with reasoning and sources, routes drafts through Slack approval, logs rep actions to LangSmith, extracts style memories from edits, and runs account-intelligence subagents in parallel.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  salesforce-lead-trigger["salesforce-lead-trigger"]
  deep-agents-orchestrator["deep-agents-orchestrator"]
  do-not-send-checker["do-not-send-checker"]
  lead-research-toolset["lead-research-toolset"]
  outbound-skill-playbook["outbound-skill-playbook"]
  slack-approval-flow["slack-approval-flow"]
  langsmith-trace-store["langsmith-trace-store"]
  rep-memory-store["rep-memory-store"]
  account-intelligence-subagents["account-intelligence-subagents"]
  langsmith-deployment-runtime["langsmith-deployment-runtime"]
  evaluation-suite["evaluation-suite"]
  salesforce-lead-trigger --> deep-agents-orchestrator
  deep-agents-orchestrator --> do-not-send-checker
  deep-agents-orchestrator --> lead-research-toolset
  deep-agents-orchestrator --> outbound-skill-playbook
  deep-agents-orchestrator --> slack-approval-flow
  deep-agents-orchestrator --> account-intelligence-subagents
  do-not-send-checker --> lead-research-toolset
  lead-research-toolset --> outbound-skill-playbook
  outbound-skill-playbook --> slack-approval-flow
  outbound-skill-playbook --> rep-memory-store
  slack-approval-flow --> langsmith-trace-store
  slack-approval-flow --> rep-memory-store
  langsmith-trace-store --> evaluation-suite
  langsmith-trace-store --> rep-memory-store
  account-intelligence-subagents --> langsmith-deployment-runtime
  account-intelligence-subagents --> deep-agents-orchestrator

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`salesforce-lead-trigger`	starts lead processing when a new Salesforce lead appears	`deep-agents-orchestrator`
`deep-agents-orchestrator`	runs the long-running GTM workflow, planning checklist, virtual filesystem, and tool orchestration	`do-not-send-checker`, `lead-research-toolset`, `outbound-skill-playbook`, `slack-approval-flow`, `account-intelligence-subagents`
`do-not-send-checker`	checks support tickets and recent teammate outreach before allowing an automated draft	`lead-research-toolset`
`lead-research-toolset`	collects CRM, Gong, LinkedIn, company, Exa, Apollo, and BigQuery context	`outbound-skill-playbook`
`outbound-skill-playbook`	conditions drafting on relationship state and defined outreach rules	`slack-approval-flow`, `rep-memory-store`
`slack-approval-flow`	presents drafts, reasoning, and send/edit/cancel controls to reps	`langsmith-trace-store`, `rep-memory-store`
`langsmith-trace-store`	records actions, traces, evals, and outcome correlations	`evaluation-suite`, `rep-memory-store`
`rep-memory-store`	stores per-rep style observations extracted from Slack edits in PostgreSQL	`outbound-skill-playbook`
`account-intelligence-subagents`	run constrained sales and deployed-engineer subagents over account data in parallel	`langsmith-deployment-runtime`, `deep-agents-orchestrator`
`langsmith-deployment-runtime`	provides horizontal scaling and durable execution for parallel subagents	`account-intelligence-subagents`
`evaluation-suite`	uses rule assertions, LLM judges, mocked APIs, and CI to catch regressions before production changes	`langsmith-trace-store`

Data flow

A Salesforce lead trigger starts the agent. It first checks whether outreach should be suppressed, then pulls Salesforce, Gong, LinkedIn, and web context, loads a relationship-aware outbound skill, drafts the email, sends a Slack DM with reasoning and approval buttons, optionally queues follow-ups, logs send/edit/cancel actions to LangSmith, and stores structured style observations in PostgreSQL for future drafts.

3 supporting quotes

“It triggers on new Salesforce leads, checks whether we should reach out, gathers context (including meeting history), and sends a Slack draft”article, stated

“The rep sees the finished draft in a Slack DM with buttons to send, edit, or cancel. They can also see the agent's reasoning”article, stated

“Those observations are stored in PostgreSQL, keyed per rep, and every future run reads them before drafting.”article, stated

Why this architecture

LangChain chose an agentic architecture because lead research is multi-step, tool-heavy, and relationship-sensitive. Deep Agents handles spiky tool outputs and planning, Slack preserves human approval, LangSmith ties actions to traces and evals, memory converts rep edits into future style guidance, and subagents let account intelligence scale while keeping tool scopes isolated.

3 supporting quotes

“The agent needed to pull from multiple sources, reason across them, and produce a personalized output. This is more than a simple LLM call can handle reliably.”article, stated

“large tool results get offloaded into a virtual filesystem automatically, so we didn't have to build our own truncation and retrieval layer.”article, stated

“Nothing is sent without an explicit rep review and approval. A single poorly timed email can undo months of relationship-building.”article, stated

Scale

Before the agent, reps spent about fifteen minutes researching before writing.
“Fifteen minutes of research before a single word was written, and no easy way to know if a teammate had already reached out yesterday.”article, stated
Lead-to-qualified-opportunity conversion rose 250% from December 2025 to March 2026.
“Lead-to-qualified-opportunity conversion rate up 250% from December 2025 to March 2026”article, stated
Pipeline dollars increased 3x in the same period.
“driving 3x more pipeline dollars in the same period”article, stated
Reps reclaimed 40 hours per month each, totaling 1,320 hours across the team.
“Sales reps reclaimed 40 hours per month each, totaling 1,320 hours across the team”article, stated
A 48-hour SLA sends silver-lead drafts automatically if a rep does not approve or decline.
“we added a 48-hour SLA for silver leads: if a rep hasn't approved or declined the draft within that window, it sends automatically.”article, stated

Alternatives considered (roads not taken)

Simple LLM call with a few tools

Rejected because: The workflow needed multi-source reasoning, state, planning, and personalized output.

“This agent required much more than a simple LLM call with a tool or two.”article, stated

Fully autonomous outbound by default

Rejected because: The launch design required explicit rep review because a mistimed email could harm relationships.

“Nothing is sent without an explicit rep review and approval. A single poorly timed email can undo months of relationship-building.”article, stated

Build custom truncation/retrieval for huge tool outputs

Rejected because: Deep Agents' virtual filesystem handled large tool results automatically.

“large tool results get offloaded into a virtual filesystem automatically, so we didn't have to build our own truncation and retrieval layer.”article, stated

Retrofitted evals after production

Rejected because: The team set up evaluations from the start to catch prompt and model-swap regressions.

“setting up evaluations from the start rather than retrofitting them later, which turned out to be critical for catching regressions”article, stated

Patterns

`multi-agent-orchestration` (orchestration)

The GTM process is treated as a long-running, multi-step agent workflow rather than a single drafting prompt.

“We built it on Deep Agents because this is a long-running, multi-step process that has to orchestrate multiple tools and large amounts of data reliably.”article, stated

`human-in-the-loop-gate` (governance)

Sales reps explicitly review and approve generated outreach before it is sent.

“Human-in-the-loop: Nothing is sent without an explicit rep review and approval.”article, stated

`observability-tracing` (operations)

Rep actions are attached to traces so the team can evaluate draft quality and regressions.

“Every rep action (send, edit, cancel) is logged to LangSmith and attached to the underlying trace so we can evaluate quality, catch regressions, and quantify what’s working.”article, stated

Tradeoffs

Keep outreach human-gated, then add limited automation for lower-priority leads after trust increased.

Chose rep approval by default with a 48-hour SLA for silver leads over fully autonomous outreach from the start.

Rationale: The case states approval was non-negotiable, then describes an added automatic send path only after refinement.

“As we've refined the agent, we added a 48-hour SLA for silver leads: if a rep hasn't approved or declined the draft within that window, it sends automatically.”article, stated

Evaluation

Representative actions are logged to LangSmith traces for quality evaluation and regression detection.

Metrics: send/edit/cancel actions