Siddhartha Borah from Brackett: teaching agents work that was never written down (Heavybit DevGuild)

The podcast discusses Bracket AI's approach to solving key challenges in automation by developing agents that can learn and execute complex, real-world business workflows. A major focus is the "last mile" problem in automation, where traditional systems fail due to missing contextual understanding - particularly the unwritten, tacit knowledge that human workers use daily. Bracket addresses this by capturing procedural knowledge through direct interaction, such as screen sharing and real-time demonstrations, allowing agents to learn not just from data but from human expertise.

The system integrates structured data (like ERP systems), unstructured data (such as emails), and human knowledge to create deterministic, repeatable workflows. Unlike exploratory AI models, Bracket's agents produce consistent, verifiable results and require only one-time corrections. The agent learning process mirrors an apprenticeship model, where workers teach by doing rather than explaining. The platform also includes advanced memory and learning mechanisms - such as confidence scoring, exception handling, periodic data "dreaming," and a business-specific knowledge graph - to ensure agents can manage unexpected inputs and improve over time. Multiple background agents support the primary agent by retrieving relevant past experiences, enhancing decision-making without overwhelming context.

• What if you could train an AI agent to handle your most time-consuming solo tasks by simply showing it how you do them?

  • Move: Record yourself completing a high-frequency, rule-heavy task (e.g., client onboarding, invoice processing) using screen + voice narration, then feed it into a learning-based automation tool like Bracket AI.
  • Why Now?: Tools that capture tacit knowledge through demonstration are now emerging from stealth (e.g., Bracket AI), making it possible for solos to offload complex workflows without coding or documentation overhead.
  • Expected Upside: Reclaim 5 - 10 hours/week within 60 days by automating one core workflow with repeatable, deterministic output - no ongoing tuning needed after initial training.

• What if you offloaded decision-making on edge cases in your software business to an AI agent that knows your historical choices and escalates only when truly stuck?

  • Move: Implement an "agent-in-the-loop" system where the AI processes routine support tickets, billing exceptions, or feature requests, using your past decisions as memory, and flags only low-confidence cases for your input.
  • Why Now?: Advances in confidence scoring and exception memory (like Bracket's deja vu and dreaming systems) now allow solo operators to delegate judgment-heavy tasks safely, with built-in escalation triggers.
  • Expected Upside: Reduce reactive work by 40% over 90 days while maintaining control - only engaging when the agent's confidence drops below a set threshold, ensuring quality and consistency.

• What if you built a personal automation agent trained exclusively by your own work patterns - no APIs, no docs, just you doing your job?

  • Move: Use a demonstration-based learning platform (e.g., Bracket AI) to train an agent over two weeks by performing real tasks while narrating decisions, letting it extract hidden heuristics (e.g., "I always double-check X before sending Y").
  • Why Now?: Apprenticeship-style AI training is now viable for individuals, eliminating the need for technical specs or process mapping - ideal for solos with messy, multi-system workflows.
  • Expected Upside: Achieve 80%+ automation of a critical workflow (e.g., reporting, data cleanup) in under 30 days, with the agent replicating your judgment, not just your clicks.

• Implement a screen-sharing and real-time observation setup to capture tacit knowledge when automating workflows, especially for processes handled by non-technical staff.
• Design automation workflows to integrate structured data (e.g., databases), unstructured data (e.g., emails), and documented human expertise to avoid the "10% context gap" that breaks automation.
• Build in a confidence scoring system for automated tasks that triggers human review when uncertainty exceeds a threshold, ensuring reliability without full manual oversight.
• Create a persistent memory layer (e.g., knowledge graph or exception log) that stores edge cases and learned rules from human interventions, enabling one-time fixes and consistent future performance.
• Adopt an apprenticeship model for training automation agents: record actual user demonstrations of tasks instead of relying on written specifications or technical translations.