AI Agents · Internal & client builds

Autonomous Multi-Tool AI Agents (Relevance.ai)

Autonomous multi-tool AI agents built on Relevance.ai — a LinkedIn outreach-personalization pipeline and a task-management assistant — that do real knowledge-work with validation and human-approval gates.

Relevance.ai — autonomous agent tool library

The problem

Personalized prospecting and task triage eat hours of manual research and clicking across apps. Automating them naively risks an LLM acting unchecked, so it needs validation and approval controls.

Approach

Each agent is decomposed into discrete, validated tools (fetch → clean → validate → generate → choose), composed into an agent that selects and invokes them, with per-tool auto-run vs human-approval modes.

What I built

• A LinkedIn personalization pipeline: scrape a prospect's posts/profile, clean and validate the data, generate candidate messages, enforce a non-salesy tone, and choose the final message
• A task-management assistant with full CRUD tools over projects, tasks, and labels, executed in the background
• Per-tool approval gating (auto-run vs human-in-the-loop) with max-auto-run limits
• Tool-to-tool chaining and custom JavaScript steps for logic the platform doesn't cover natively
• Multi-channel action tools (Slack, Gmail, Airtable, voice) for an SDR-style agent

Result

Shipped autonomous agents that turn multi-step research-and-writing and task workflows into hands-off, guardrailed pipelines — with human approval where it matters.