Apollo.io rebuilt the AI assistant inside its sales platform from a supervisor that routed work to fixed subagents into a single planning loop that picks its own steps, according to a conference talk by Aniruddha Laud and Anshul Pahwa of Apollo.io, published on September 29, 2026.
TL;DR: Apollo.io's first assistant used a supervisor that handed each request to a dedicated subagent for prospecting, enrichment or research. Every new capability needed a new subagent, and users faced several confirmations per request. The rebuild uses one loop with a planning tool, a skills library, a virtual file system and short-lived subagents, plus a setting where customers choose when the agent stops to ask. The speakers report that latency went up, but 3-day and 7-day retention rose 31% and 38% (13:56), and users were 2.3x as likely to book meetings (14:07). The talk does not state the comparison period or the method.
| Fact | Detail |
|---|---|
| Source | View source |
| Source type | YouTube talk (LangChain Interrupt NYC), 16 min |
| Source published | 2026-09-29 |
| Speakers | Aniruddha Laud, Director of Engineering. Anshul Pahwa, Senior Manager, Engineering. Both at Apollo.io |
| Industry | Sales software |
| Function | Sales |
| Techniques | Single planning loop, skills library, virtual file system, subagents, approval settings, evals |
| Tools named | LangGraph, Deep Agents, LangSmith, Slack |
Independent summary of public material. Apollo.io is not affiliated with Taskade.
The system Apollo.io built
Apollo.io sells a platform for sales and go-to-market work, from prospecting and data enrichment to email and closing deals. Aniruddha Laud reports more than 40,000 paying customers (01:34) and millions of users (01:54), so a new feature must work at full scale on day one.
Customers gave three kinds of feedback (02:40) about the product before the agent:
- The product grew too wide, so each new feature meant a new workflow to learn.
- Research on companies and people took a lot of manual time.
- Users switched often between tasks that need very different thinking.
The team answered with an assistant that answers questions and also acts for the user. Work started a little more than a year before the talk (04:41). The first version shipped and reached general availability earlier in 2026 (04:43).
Architecture of the system
The first version used a supervisor with subagents. The supervisor chose which subagent to call, such as a prospecting agent or an enrichment agent. Each subagent ran one fixed slice of a workflow. Anshul Pahwa reports these problems with it:
| Problem | What the speakers describe |
|---|---|
| Engineering cost | A new subagent reached about 20% quality fast, but 80% took many evals and much tuning (07:32) |
| User cost | One question triggered up to five confirmations, and users dropped off (07:53) |
| Context | Each subagent passed context to the next, so long conversations grew too large |
| Shipping cost | Every new capability went through the core team, which lacked capacity |
The rebuild has a different goal: let the model decide the steps, and make new capabilities fast to add. The new design is one loop with these parts:
- A planning tool that decides which skills to use.
- A skills library. Each new capability is one file that describes what it does, with no extra routing.
- A virtual file system that holds context so that it does not overflow.
- Subagents that the loop spawns to isolate a piece of work.
- A governance layer where customers set when the agent stops to ask. For example, a customer can require approval before the agent spends more than 20 credits on a set of actions (10:29).
The team then built a meta skill that turns a natural-language description into a new skill for this design. The speakers chose their framework for model neutrality, openness and their existing tracing and eval setup. They also evaluated a coding-agent SDK. The move from the old design to general availability took two to three months (05:15).
Results the source reports
- Anshul Pahwa reports that latency rose significantly, because the agent now plans its own path (13:38).
- Pahwa reports a 31% lift in 3-day retention and a 38% lift in 7-day retention (13:56).
- Pahwa reports that customers were 2.3x as likely to book meetings (14:07).
- Pahwa reports that more than 50% of customers rely on the assistant and agent features (14:13).
- Pahwa reports that the meta skill raised development velocity by 80% to 85% (12:46), and that a new capability reaches a first version in two to three days (12:55).
- On latency, Pahwa says customers "rage quit if you're not doing what you're promising" (13:42).
Critical assessment
The speakers do not give the period, the comparison group or the method behind the retention and meeting figures, and the talk does not rule out other product changes in the same period. All numbers come from Apollo.io itself, and no outside party checked them. The talk took place at a conference run by the vendor of the framework that the team chose. The problems with the first design are specific, and they are the most reusable part of the source. Apollo.io names Slack, MCP and a command-line interface as next surfaces, plus autonomous agents that run in the background.
Build this in Taskade
- Build an AI agent with web search and persistent memory that researches accounts. Start at AI agents.
- Start from a template in sales agents.
- Run Apollo steps inside a workflow with Apollo automations.
- Pull Apollo data into your workspace with the Apollo integration, one of 100+ integrations.