Three Kinds of Software Survive: Tasklet's Andrew Lee on Competing to be a Horizontal Platform

Episode Summary

Executive Summary: Andrew Lee says Tasklet has rewritten nearly its entire stack to support a general-purpose, synchronous agent built around file-system-based memory, heavy summarization, and computer-use-first execution. He argues the AI software market is converging, so Tasklet must become a neutral horizontal platform spanning multiple model providers while balancing rising token costs, model quality changes, and competition from vendors like Anthropic and OpenAI.

Main Topics: Total product and stack rebuild (Priority: 5/5): Tasklet evolved from a workflow automation tool into a broad, synchronous agent platform. This required rewriting the UI, agent architecture, integrations, computer-use layer, and context management. Context management, summarization, and caching (Priority: 5/5): Tasklet moved history into a file system and uses layered summaries with decreasing fidelity over time to preserve long-running agent memory while controlling token spend and cache efficiency. Model-provider strategy and the Anthropic/OpenAI transition (Priority: 5/5): Andrew explains why Tasklet stayed Anthropic-first, how newer models changed what is possible, why 4.7 was not a default due to cost, and why OpenAI, Google, and open-source models are becoming viable options. Competition, pricing pressure, and horizontal-platform thesis (Priority: 5/5): Anthropic is both a key supplier and a direct competitor through consumer plans that subsidize usage. Tasklet’s response is to become a neutral platform that can route work to many models and survive as one of a few horizontal winners. Harness/mecha-suit philosophy and model-harness tradeoffs (Priority: 4/5): Lee reframes the 'harness' as a mecha suit: low-level primitives, persistence, browser/shell/file-system access, oversight, and tooling that let models act in the world. He argues model and harness improvements multiply each other. Shared brain, organizational context, and multi-agent memory (Priority: 4/5): Tasklet is adding organization/workspace/agent-level context, shared connections, and eventually cross-agent memory so companies can use a common situational awareness layer across agents. Future software landscape and product categories (Priority: 4/5): Andrew predicts software will consolidate into horizontal platforms, headless API-first companies, and solutions businesses. He sees products like Salesforce as vulnerable and emphasizes durability, rollback, and approvals as key to trust.

Key Arguments: Tasklet had to be rebuilt because users wanted one continuous conversational agent, not separate workflow and chat products. Long-term chat history cannot be naively fed into models; the durable solution is file-system memory plus structured summarization and selective retrieval. Caching remains crucial because file-system-backed agents require many more tool calls, so per-run and provider-specific caching materially affects cost. Anthropic models enabled Tasklet early, but Anthropic’s consumer max plans create direct competitive pressure by subsidizing token consumption at roughly a much lower effective price. Tasklet is moving from model dependency to a multi-model abstraction layer so customers can benefit from whichever frontier model is best for a task without having to manage model bets themselves. Model and harness improvements are orthogonal and multiplicative; the best production systems require both strong models and strong orchestration, oversight, and memory. The most durable software winners in the AI era will likely be a small number of horizontal platforms, API-first infrastructure companies, and outcome-selling services. Shared organizational memory, permissions, and rollback/approval workflows will be essential for reliable enterprise adoption because agents will act on behalf of teams over long periods. The biggest competitive risk is that model vendors increasingly ship products that look like Tasklet, but Tasklet’s differentiator is being neutral, persistent, and cross-model. Tasklet’s internal and user-facing economics depend on controlling token spend, which influences model choice, caching policy, and architecture decisions.

Data Points: Time since last interview: About 6 months - Andrew describes what has changed since the prior conversation. Model quality/cost change for Opus: 15 to 5 - He says Opus price dropped from 15 to 5, which made it much more viable for Tasklet. Relative API vs consumer token access: Approximately 5:1 - Andrew estimates Claude Max customers can consume about five times as many tokens as Tasklet can buy through the API at the same price. Cost impact of Claude 4.7 tokenizer changes: ~30% higher cost - Tasklet chose not to default to 4.7 because tokenizer changes would raise costs significantly. Caching window on Anthropic: 5 minutes - Tasklet currently relies on short-lived Anthropic caching for active sessions and turns. User churn destination: 80% - Andrew says about 80% of users who leave Tasklet go to an Anthropic product. Internal token spend vs payroll: 5% to 10% - He estimates Tasklet’s internal token spend is roughly this share of payroll. Token spend products used internally: 3 major products - He cites Claude, Codex, and Tasklet as major internal token-consuming tools.

Pivotal Quotes: "speed is the only moat" — Andrew Lee: A core operating principle that motivates Tasklet’s frequent rewrites and rapid adaptation to model changes. "what if the files are the agent?" — Andrew Lee: Describes Tasklet’s shift from putting full history in context to storing history in a file system and retrieving it selectively. "we're going to give you Anthropic models and OpenAI models and Google's and all the open source models" — Andrew Lee: Explains Tasklet’s strategy to become a neutral horizontal platform rather than a single-model product.

Implications: Builders should expect AI products to become multi-model, memory-rich, and heavily permissioned. Winning apps will likely be platforms with durable context, rollback, and cost control—not just front ends on top of one model.