How Clay Scales GTM Agents

Business context & objective (GTM agents at “production scale”)

Jeff Barr (Head of AI at Clay) frames go-to-market (GTM) as requiring continuous iteration because “no creative advantage lasts forever.” For example:

• Cold email deliverability can degrade over time.
• GPT-era copy can sound human, but still may not solve product-market fit.

Goal: scale from running single agents effectively to running many agents across your entire addressable market—reliably, cost-effectively, and at high quality.

Company positioning / operating model (what Clay does)

Clay positions itself as GTM infrastructure, enabling the end-to-end agent workflow for GTM execution:

• Builds company/person lists from GTM data
• Enriches lists via integrations and AI agents
• Orchestrates lists into workflows such as:
  • CRM enrichment
  • outbound campaigns

Scale claims / capacity:

• 350M+ go-to-market agent runs per month
• 40M+ companies and 900M contacts in proprietary datasets
• Agent research running over trillions of tokens per week

Frameworks / playbooks mentioned

“Go-to-market alpha” (finance analogy)

“GTM alpha” refers to durable outperformance driven by:

• Better audiences
• Better timing
• Better signals
• Better positioning

Three levels to reach GTM alpha

1. Individual AI literacy & enablement Equip sellers with tools like ChatGPT/Claude for tasks such as:

   • call analysis
   • outbound copywriting

2. Centralized deployment Use AI at the workspace/workflow level (e.g., auto-generating post-call notes via Claude).

3. Non-transferable advantage Create strategies competitors can’t easily copy (e.g., unique outbound plays built on niche signals).

Account selection loop (scan → score → act → learn/iterate)

A recurring operational loop:

• Scan the entire addressable market
• Layer signals (news, fundraising, and bespoke signals such as GitHub stars, hiring intent, etc.)
• Score accounts to determine when to reach out
• Act at the right time
• Learn from outcomes and iterate plays over time

Key business insights & recommendations (actionable)

• Avoid getting stuck at Level 1 (writing better emails). Even excellent copy won’t matter if prospects aren’t already in an “interested-to-buy” state.

• Shift GTM effort toward high-leverage targeting by:

  • using agents to identify accounts with intent signals
  • reaching out at the optimal moment
• Treat scaling GTM agents as an engineering problem, not only “marketing automation.”
• Implement a measurable feedback loop / observability:
  • use offline + online evaluation to tune the agent harness for GTM outcomes.

Scaling challenges & operational solutions (Claygent deployment at production scale)

Clay describes four core “production scale” challenges and corresponding engineering approaches.

1. Infrastructure reliability

• Problem: agents spend significant time waiting (browsers, APIs, inference). Running on AWS Lambda became prohibitively expensive due to wall-time billing.

• Solution: move to ECS, while re-architecting reliability with durable workflows:

  • queues
  • checkpointing at periodic steps
  • tools mentioned: LangGraph / LangSmith

2. Rate limits & throughput

• Problem: workloads are spiky; you must keep inference “productive” despite downstream rate limits.
• Solution: implement adaptive backpressure/throttling inspired by TCP/IP congestion control:
  • send as much traffic as allowed
  • progressively dial back when rate limits occur
• Reported impact: internal experiments showed 4x to 10x more throughput than naive systems.
• Additional requirement: fairness across customers
  • prevent one customer’s large runs from crowding out others (e.g., millions vs tens of onboarding accounts)

3. Cost

• Cost lever #1: caching strategies
  • caching within the agent harness reduced repeated work
  • for providers like Anthropic, reported savings: up to 70%
• Cost lever #2: bound retries and tool calls
  • constrain execution steps/research time
  • must be paired with evaluation to ensure quality doesn’t degrade
• Cost framing: measure cost tied to quality and outcomes, not only token spend.

4. Quality

• Root drivers:
  • Great context (high-quality web access + proprietary GTM datasets)
  • A tuned harness for GTM use cases, including:
    • offline evals + online evals
    • tools mentioned: LangSmith to understand optimization and failure modes
• “Quality is also a product problem”
  • Clay built an agent builder to test and iterate before market-scale runs.

Metrics / KPIs and performance targets mentioned

• Throughput scaling (engineering KPI): 4–10x throughput improvement from backpressure vs naive throttling
• Cost KPIs: up to 70% cost savings via caching (Anthropic example)
• Capacity / volume (operational KPI / scale):
  • 350M+ runs/month
  • trillions of tokens/week
• Revenue/cash KPIs: none explicitly provided (e.g., CAC/LTV/churn).

“What’s next” product direction (audiences + agent memory)

• New product: Audiences (built over ~6 months)
  • aggregates GTM data into a single place from sources such as:
    • Snowflake, Salesforce, Gong, call recordings
  • adds third-party signals like:
    • fundraising announcements
    • news
• Feeds into agent memory / “go-to-market intelligence”
  • agents recommend plays based on what has worked before
  • described as a flywheel:
    • test outcomes → learn → improve plays over time

Additional infrastructure mentioned:

• virtual file systems for reasoning over context
• sandboxes for safer experimentation/testing

Source / presenters

• Jeff Barr, Head of AI at Clay