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

What It Covers

Tasklet CEO Andrew Lee details a complete architectural rebuild over six months—shifting from workflow automation to a general-purpose agent platform using file system-based context management, multi-provider model support, and organizational memory features. Lee also maps out which three categories of software companies survive the AI transition and explains why Anthropic is simultaneously Tasklet's best partner and most threatening competitor.

Key Questions Answered

• •File System as Agent Memory: Rather than feeding entire chat histories into LLM context windows, Tasklet stores all history in a persistent file system and sends only compressed hints to the model. Context is compressed in time-bucketed layers—recent turns retain full fidelity including thinking blocks and tool call responses, while older turns are progressively stripped down to LLM-summarized abstractions. This approach scales agent memory far beyond context window limits while keeping token costs manageable across thousands of trigger-based runs.
• •Multi-Resolution Summarization for Long-Running Agents: Tasklet's compaction system uses decreasing fidelity as conversation history ages: tool call responses are truncated first, then arguments collapsed, then assistant messages shrunk, then LLM-based summarization applied in discrete buckets. Buckets are designed to minimize cache prefix disruption. This architecture is specifically built for agents that fire 10,000+ times annually on recurring triggers—a fundamentally different optimization target than single-session coding agents that reset context between runs.
• •Supplier-Competitor Pricing Gap: Anthropic's Claude Max subscription delivers an estimated five times more tokens per dollar than Tasklet can purchase via the API at equivalent cost. Approximately 80% of Tasklet churned users migrate directly to an Anthropic product. This structural pricing disadvantage is the primary driver pushing Tasklet toward multi-provider neutrality—positioning itself as a platform that benefits from all model improvements rather than being dependent on one supplier's pricing decisions.
• •Three Software Categories That Survive AI: Lee identifies the only durable software business models as: horizontal agent platforms (very few winners, likely two to three per category), headless API-first companies like Stripe where compliance complexity justifies existence without a UI, and solutions companies that sell outcomes rather than software—exemplified by charging per resolved customer service ticket. Traditional SaaS products with bespoke UIs, including Salesforce, face structural obsolescence as general-purpose agents generate custom interfaces on demand.
• •Model Selection and Cost Trade-offs: Tasklet chose not to deploy Claude 4.7 as a default model because it increased token costs approximately 30% without meaningful performance gains for iterative knowledge work automation—the core Tasklet use case. GPT-5.5 now matches Opus 4.6 for most Tasklet workflows and will launch as an alternative. Kimi's latest model benchmarks near Haiku performance at lower cost, making it a viable option for cost-optimized agent tasks within a multi-provider architecture.
• •Generative UI Eliminates Vertical SaaS: Tasklet's Instant Apps feature lets users generate fully functional, data-connected interfaces in a single prompt—replacing dedicated dashboard tools, BI platforms, and workflow UIs. Internal teams now generate BigQuery explorer dashboards on demand rather than maintaining separate tooling. This validates Lee's thesis that vertical AI-embedded products have a limited shelf life: a general-purpose agent with generative UI can replicate most SaaS interfaces on demand, collapsing the differentiation that justified standalone products.
• •Organizational Context Hierarchy for Teams: Tasklet is building a three-tier shared context architecture: organization level for company mission and values, workspace level for team OKRs and brand guidelines, and agent level for workflow-specific instructions. Shared connections already allow team leads to configure API credentials once, making them available to all team members instantly. Planned additions include cross-agent memory, shared file systems, and skill libraries—enabling agents to retain institutional knowledge across the entire organization rather than siloing context per agent.