YC Root Access

How to Build an Internal AI Agent That Evolves Itself

AnswerThis builds AI agents for evidence-based scientific workflows and has scaled past $2 million in ARR with just two full-time employees — largely because they built an internal AI ops agent that processes over 100 emails a day, closes support tickets, and updates their CRM automatically. In this recent batch talk, founder Ayush Garg breaks down the architecture of a self-extending agent that builds its own tools when it encounters tasks it can't handle yet, how his non-technical co-founder trains the agent by giving it feedback in Slack, and the three types of memory — factual, behavioral, and procedural — that any founder can copy to build an internal agent for their own business.

May 19, 2026·5m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   Why an internal AI ops agent is a leverage multiplier for tiny teams

   Ayush introduces AnswerThis and explains how a two-person founding team reached $2M+ ARR by offloading operational load to an internal AI agent. The agent reduces context switching and founder time by centralizing work and answers that would otherwise live across many tools.

2. 2

   What the agent actually does day-to-day (email, support, CRM, feedback)

   He lists concrete workloads the agent handles across communication and customer workflows. The emphasis is on real operational throughput and breadth of responsibilities rather than demos.

3. 3

   Instantly queryable business context: ask questions any time

   Beyond task execution, the agent becomes a conversational interface to company operations. Founders can ask ad-hoc questions and get answers without manually checking multiple systems.

4. 4

   The differentiator: a self-extending agent that creates new tools

   Ayush stresses the key feature: the agent isn’t limited to a fixed toolset. When it encounters a repeated task it can’t perform, it delegates to a coding sub-agent to build a new tool that persists for future use.

5. 5

   Core harness architecture: Claude Code CLI + Python task queue

   He outlines the foundational setup: a coding-capable CLI agent (Claude Code) wrapped in a thin Python harness. Messages from Slack/email flow into a task queue that the agent processes iteratively.

6. 6

   Teaching business-specific logic via read-only codebase + database snapshots

   To give the agent domain knowledge that founders usually carry, they provide a read-only copy of the database and codebase. A cron job refreshes this context on each release so the agent can answer support and product questions by inspecting the source of truth.

7. 7

   Tooling layer: startup app CLIs plus a coding agent that can edit the agent itself

   The agent is equipped with CLIs for core SaaS tools (Intercom, Fathom, Stripe, etc.). Crucially, it also has access to a general coding agent as a CLI that can modify the agent’s own code, enabling tool creation on demand.

8. 8

   Proof of self-evolution: growing to 45+ CLIs and creating monitoring automatically

   Ayush shares how the agent expanded from a basic skeleton to a mature system by authoring tools itself. A concrete example: after being asked to monitor landing pages for ad uptime, it created a cron job to do it.

9. 9

   Editable personality & memory: instructions.md that the agent updates every turn

   A key mechanism for behavioral improvement is an instructions.md file loaded at every agent turn. The agent can edit this file, allowing feedback to directly reshape future behavior without requiring engineering intervention.

10. 10

    Operational example: non-technical support feedback that permanently fixes mistakes

    He describes how his non-technical co-founder corrected a class of customer support mistakes by messaging the agent in Slack. The agent updated its instructions/tooling so the same error pattern stopped recurring.

11. 11

    The three-memory model for effective internal agents

    Ayush abstracts the system into three memory types that make internal agents work well: factual, behavioral, and procedural. Each maps to a concrete implementation choice in their architecture.

12. 12

    Blueprint to replicate the setup (minimal steps)

    He closes with a practical recipe for building a similar agent: use a coding-capable CLI harness, provide read-only access to sources of truth, seed it with basic CLIs, add a coding sub-agent, and maintain an editable instruction file. Connect it to Slack/email and let it grow capabilities over time.