Three Kinds of Software Survive: Tasklet's Andrew Lee on Competing to be a Horizontal Platform
Summary
Tasklet is betting that the agent transition leaves only three viable software archetypes: a few horizontal platforms, API-first companies, and outcome-selling solutions firms. Its ambition is to become “the AI agent platform that replaces your SaaS products for knowledge workers,” while Stripe-like infrastructure survives behind APIs and Nathan cited Fin’s model of $0.99 per customer-service ticket resolved as an example of selling outcomes rather than visible software.
Tasklet’s greatest supplier is also its clearest competitive threat. Andrew Lee says roughly 80% of departing users move to an Anthropic product, usually because they already pay for Claude Max; Nathan and Lee guess that Max may deliver “five to one or maybe even more” tokens per dollar than Tasklet can buy through the API. That subsidy distorts customer expectations and keeps Tasklet near razor-thin margins. Separately, Opus 4.7’s roughly 30% higher tokenizer-driven cost helped make it an optional model rather than the default.
The six-month rewrite replaced chat history with a cache-aware file-system architecture built for agents that may run 10,000 times. Tasklet now stores the complete history in files and sends the model hints plus a fixed-length, decreasing-fidelity summary: recent turns retain tool calls and thinking, while older material progressively loses responses, arguments, and detail before reaching LLM summarization. “What if the files are the agent?” is the central design call; the remaining weakness is that agents still sometimes forget and compaction still costs heavily.
“Always bet on the models” held up, but Tasklet is shifting from Claude maximalism to model-neutral orchestration as credible alternatives arrive. Claude 4.5 unlocked better computer use, Opus’s price move from $15 to $5 widened access, and 4.6 improved computer use and code generation enough to support Instant Apps. Lee says GPT-4.5 has become very good for Tasklet’s use case, can navigate its harness well, and gives Opus 4.6 “a run for its money”; Tasklet has signed a deal with OpenAI and expects Anthropic, OpenAI, Google, and open-source models to offer useful cost-performance choices. Kimi and DeepSeek are among the models it has tested.
Tasklet’s defensible wedge is not exclusive capability but the ergonomics and governance of persistent corporate automation. Lee concedes that “everyone is building the same thing” and that almost any general agent can code or perform knowledge work; Tasklet instead optimizes for cloud-hosted, 24/7 workflows with shared ownership, audit logs, guardrails, cost controls, and memory across thousands of triggers. Its enterprise pitch is therefore, “A bet on us is a bet on everybody,” with Tasklet serving as a neutral model and cost arbiter.
Generative UI is arriving fast enough to erase large portions of application-layer differentiation. Tasklet’s March Instant Apps release can generate connected dashboards or even a functioning email interface from one prompt; internally, the team now asks Tasklet for interactive pricing-analysis tools instead of building them in BigQuery or conventional dashboard software. Lee consequently thinks “Salesforce is in real trouble”: agents make schemas easier to recreate, data easier to move, and accumulated application code less valuable, though he expects a smaller Salesforce rather than extinction.
Better models do not eliminate the commercial value of the harness; they move that value toward cost, reliability, permissions, and reversible execution. Lee prefers “mecha suit” to harness: storage, compute, APIs, persistent context, and oversight multiply the model’s usefulness even if the intelligence advantage lasts only six months. The roadmap includes the ability to “roll back the world,” require approval only for consequential actions, and generate testable migration scripts rather than passing records through an LLM; meanwhile, Tasklet’s own internal token spend is estimated at 5–10% of payroll.
Deep dive
1. Tasklet rebuilt everything because workflow automation was too narrow
Lee could identify “nothing substantial” that survived the prior six months: visual design, application structure, connections, computer use, agent core, context management, and compaction were all rebuilt. The governing mantra remained “speed is the only moat,” but almost every assumption beneath the product changed.
The October product asked users to describe workflows and let Tasklet execute them. Immediate feedback changed the scope: once customers had given an agent their context and access to their systems, they did not want it limited to asynchronous workflows; they also wanted to converse with that same agent synchronously.
That demand created a deceptively hard product requirement: one long, linear conversation containing both interactive work and recurring automations. Feeding every prior run back into the model would eventually mean sending millions of irrelevant tokens whenever a trigger fired, making the desired experience technically and economically untenable.
Computer use moved from a Windows—and later Linux—machine “tacked on” as an afterthought to the product’s critical path. Agents now routinely execute shell commands, manipulate files and databases, and retain state in headless and browser VMs across runs; if computer use fails today, Lee says, “everything goes down.”
2. The file system became the agent’s durable memory
Tasklet’s architectural inversion was, “What if instead of the history being the thing we send to the LLM, what if the history is in the file system? What if the files are the agent?” The prompt becomes a compact map of what exists and what the model should inspect, expanding practical memory from context-window scale to file-system scale.
The active turn stays at high fidelity, often including thinking blocks, tool arguments, tool responses, files, and the final answer. Older turns progressively lose thinking, then tool responses and arguments are truncated or removed, calls are collapsed, assistant messages shrink, and only the oldest material receives LLM-generated summarization.
Compression happens in aging buckets so Tasklet does not constantly invalidate cached prefixes. Buckets accumulate slowly, then shrink after crossing thresholds; the operating assumption is that recent events matter more, while the agent can search the complete file-system history when an older detail becomes relevant.
Maintenance is incremental after every interaction and can occur inside a single long run. Tasklet persists the result because repeatedly summarizing old material would itself consume substantial tokens; Lee’s hedge remains important: the system “generally worked,” but users still report forgotten details and it remains expensive.
3. Caching grew more important as context moved out of the prompt
File-system context reduces the history sent by default but increases the number of tool calls needed to retrieve information, making cache efficiency more—not less—important. Tasklet’s compression scheme therefore optimizes both relevance and prefix stability rather than treating summarization as an isolated memory problem.
Anthropic caching currently lasts five minutes, which usually captures one active session or a single agent run but not the next scheduled trigger. Lee considers that acceptable because most customer automations run every few hours or once a day, not every half-hour.
OpenAI’s primitive is materially different: it automatically caches any prefix for 24 hours. Anthropic requires explicit management and allows only four cache points per call, so Tasklet needs provider-specific translation even while trying to keep the logical context identical across models.
Today, caching is effectively per agent; Tasklet receives no material reuse across agents, organizations, or users. Lee would not disclose the planned implementation, but sees substantial savings in safely sharing cacheable material at those broader levels.
4. Model gains opened products, but cost still decides deployment
Claude 4 was capable enough to start Tasklet, while 4.5 was the first major unlock for computer use and for navigating connections and tool activation. A December reduction in Opus pricing “from 15 to five” was equally consequential because Tasklet could move beyond relying mainly on Sonnet.
Lee characterizes Claude 4.6 as a solid incremental improvement in headless and browser use, plus code generation; that improvement enabled Instant Apps. Opus 4.7 is better at one-shot coding and long projects, but did not produce a comparable jump for Tasklet’s iterative knowledge-work workloads.
The 4.7 tokenizer change raised Tasklet’s measured cost by about 30%. Because those costs ultimately reach users, Tasklet declined to make 4.7 the recommended default and instead plans an advanced option clearly marked as more expensive—an unusually direct example of model economics overriding benchmark leadership.
Anthropic initially won because alternative models could not reliably navigate Tasklet’s discovery, connection activation, and context-management harness. That constraint has now broken: Lee says GPT-4.5 has gotten very good, is a huge step up for his use case over 5.4, handles the same harness well, and gives Opus 4.6 “a run for its money.”
5. Tasklet is turning single-vendor dependence into neutral-platform leverage
Lee carefully holds both sides of the Anthropic relationship: its models made Tasklet possible, its team provides early access and responsive support, and it takes Tasklet’s feedback seriously. Yet approximately 80% of users who turn off Tasklet subsequently choose an Anthropic product.
The leading cancellation reason is not necessarily capability but an existing Claude Max subscription. Every model release helps Tasklet, while every Max-plan enhancement makes selling Tasklet harder; Lee says Anthropic’s subsidized pricing has created “distorted expectations” about how much model usage a third party can provide.
Neither speaker knows the exact Max-to-API advantage, but Labenz’s intuitive estimate was five-to-one and Lee replied, “That would be my guess too—like five to one or maybe even more.” Tasklet must explain that it is operating at “pretty razor-thin margins” while competing against the supplier setting the reference price.
The strategic answer is neutrality: “A bet on us is not a bet on Anthropic or OpenAI or anyone else. A bet on us is a bet on everybody.” Labenz called the move from Claude maximalism to multi-model abstraction unusually well timed; Lee’s unhedged response was that this had “very much” been the plan.
6. Persistent corporate automation is Tasklet’s chosen wedge
Lee’s broad concession is that “everyone is building the same thing.” As models gain general tools—files, browsers, computers, and code—Claude Code, Codex, Tasklet, and other agents can increasingly perform both coding and non-coding work; differentiation moves toward optimization choices and workflow ergonomics.
Tasklet can connect to GitHub, write code, and generate pull requests; Lee even drafts marketing content inside Tasklet and lets it create the PR. He nevertheless expects a dedicated coding harness to remain smarter, cheaper, and better presented for heavy engineering work.
Tasklet instead optimizes for “24/7 automation of knowledge work for companies,” particularly work owned by the organization rather than an individual. Corporate invoicing cannot stop because someone closes a laptop or trips over a Mac Mini’s power cord; it needs cloud execution, shared management, auditability, guardrails, and cost controls.
Its memory system reflects that market: an email-triggered agent might run 10,000 times in a year and must still retain relevant early instructions. The context resets that make sense inside a bounded coding session do not fit an agent continuously processing a company inbox.
7. A mecha suit multiplies models after intelligence stops being scarce
Labenz questioned whether “harness” had become anachronistic: the work increasingly broadens a model’s world rather than restraining an unruly animal. Lee preferred “a mecha suit”—the system supplies storage, compute, APIs, user communication, memory, and control so the underlying model can act.
Users often assume the screen maps almost directly to one model call, but Lee says the translation layer is becoming much more elaborate and could grow 10 times more complex. He expects breakthroughs in memory, oversight, and tool connection to expand capabilities even while the foundational model APIs converge.
Labenz’s pushback—worth keeping—was that the capability lead from the best harness over a minimal one appears to be shrinking. Lee agreed that a current model with a poor harness can beat a year-old model with a great one, but called model and harness improvements “multiplicative” and “orthogonal”: even if engineering buys only six months, “it’s six months.”
Intelligence is not the only production metric. Once an agent is smart enough to order lunch every day, further reasoning gains may add little; cost, latency, reliability, observability, and approvals dominate. Lee pointed to Anthropic’s supervisor-agent approach, in which a smaller model can call a larger one, as a harness-level route to near-frontier performance at far lower cost.
8. Major labs are converging while new architectures could reset the race
Tasklet can test models quickly, but production support is slower because thinking blocks, prompts, caching, and failure modes vary. Its initial filter is candidly “mostly vibes”: GLM has been tested, while Google, Kimi, DeepSeek, and OpenAI appeared close enough to the frontier to justify deeper work.
User demand is an important external test. Lee recalled dismissing customers who said his earlier product was on an obsolete GPT model shortly after 3.5 appeared; “turns out they were totally right.” By contrast, he has not yet seen meaningful customer insistence that Tasklet add Grok, though he will not rule it out.
Among major labs, Lee sees imitation-driven convergence: his deliberately “flippant” read is that Opus 4.7 adopted some Codex-like precision, while OpenAI improved Codex after watching Claude Code; 5.5’s longer general-purpose agent tool calling continues the exchange. A radically different approach such as JEPA might still “shake the snow globe,” but he treats that as unresolved.
Model character may still diverge. Labenz relayed Andon Labs’ finding that GPT-5.5 ran its businesses “clean,” while Opus 4.6 and 4.7 could be “ruthless”; Lee’s anecdotal experience is that Anthropic feels more creative, empathetic, and human, while OpenAI feels clinical. Tasklet has not received reports of its agents acting unethically.
9. OpenAI adds both supply diversity and a second platform threat
Tasklet has signed an OpenAI deal and expects a broad model menu within months. Lee remains likely to recommend Anthropic for many workloads, but expects Anthropic, OpenAI, Google, and open-source models to create useful cost-performance tiers; Kimi and DeepSeek are among the models Tasklet has tested.
Supporting many providers argues for the smallest possible bespoke layer. Tasklet wants an agent’s persisted state to survive model switching without translation problems, so it prefers common tools and prompts, with modular model-specific adjustments only where APIs or behavior make them unavoidable.
OpenAI’s apparent willingness to let third-party tools use a customer’s core account could eventually allow Tasklet users to bring their own tokens. Lee would integrate that if it becomes popular and durable; he does not regard Tasklet as merely a token reseller, and sees account-based access as potentially useful onboarding.
The competitive concern increased when OpenAI killed Sora and reportedly refocused on business productivity. AgentKit had not looked like its “A game,” but Codex’s rapid move from also-ran to possible category leader shows what focused execution could do. Lee’s balancing evidence: no customer has yet told him they left Tasklet for an OpenAI product.
10. Shared organizational context is becoming the product’s connective tissue
Tasklet already has unannounced organization and workspace foundations visible in settings. Lee described a hierarchy: organization context captures the company, mission, and values; workspaces hold team resources, quarterly OKRs, processes, files, and brand voice; individual agents retain workflow plans, uploads, and conversation-specific instructions.
Shared connections are the first live workspace-level context. A technical lead can configure API keys, headers, and service access once, then make those connections available to teammates; new employees can begin using agents without locating credentials or rebuilding integrations.
The roadmap adds shared skills, cross-agent memory, and a native shared file system. If a user explains a durable fact to one agent, another should be able to retrieve it; that is the “shared brain” needed to keep synchronous assistance and recurring automation inside one coherent system.
Lee acknowledged another product that he said was called “shared brain” and offered a clearly labeled hunch: it may be farther ahead on the “brain side,” while Tasklet’s agents are stronger. His goal is to catch and surpass its context layer without surrendering the agent lead.
11. Instant Apps demonstrate why application interfaces have a shelf life
Tasklet’s strategy began with a threat to Shortwave, its still-operating AI email client. Lee realized a general agent would soon answer “show me my inbox” by generating the email interface itself; an agent embedded in a carefully built application UI therefore had a shelf life—probably much less than 10 years.
The first escape was a general workflow agent, but customers rejected separating recurring work from everyday assistance because the systems would need duplicate context. Tasklet consequently broadened again, reinforcing Lee’s conclusion that each supposedly defensible application boundary is being absorbed by a more general product.
Instant Apps, launched in March, generates a UI connected to data from any available integration in one prompt. Tasklet’s own team now requests an exploratory dashboard to model pricing changes, complete with thresholds and toggles, rather than opening BigQuery or configuring a conventional dashboard tool.
The original Shortwave fear is already testable: Tasklet can generate a working email UI today. It is not yet as good as Shortwave, but Lee thinks the gap will close quickly; “the high end of these things” arrived substantially faster than his team expected.
12. Only three software forms survive if general agents absorb SaaS
Tasklet’s end state is one horizontal agent replacing the knowledge worker’s circuit among Word, Notion, Linear, and other tabs. Data arrives through APIs, bespoke analysis becomes generated code, and the interface is produced on demand—“the best harness” becoming intelligent across essentially everything.
Horizontal platforms are the first surviving class, and Lee expects very few winners because users will not maintain the same context and connections repeatedly. A customer might tolerate one platform for knowledge work, one for coding, and perhaps one for personal use—not thousands of applications with separate embedded agents.
Headless infrastructure is the second class. Payments remain complicated, regulated, and consequential, so Stripe may endure even if nobody visits its dashboard; its durable product becomes the API. The third class sells solutions or outcomes, including AI-enabled lawyers and real-estate agents whose underlying software may be invisible. Nathan earlier cited Fin’s $0.99-per-resolution service model as an example.
Salesforce sits in the vulnerable middle. Lee’s call is categorical—“Salesforce is in real trouble”—because much of its accumulated code is obsolete, competing schemas are easier to generate, and agents reduce system-of-record lock-in by moving data more easily. He does not predict death, but “a much smaller Salesforce.”
13. Trust, unit economics, and infrastructure now matter more than demos
Labenz’s personal agent once deleted a Slack export that had taken about four rate-limited days to assemble, crystallizing the reliability requirement. Lee wants versioned file systems and action logs that let users “roll back the world”; undoing files is straightforward, but reversing external API actions may require keeping logs of what happened.
Permissioning should be consequential and ergonomic, not universally blocking. An email agent might freely read, search, and draft, yet require approval before sending; Tasklet could push a notification when the proposed message reaches that irreversible boundary.
For high-reliability migrations, Lee rejects passing records through model context and trusting reproduction. The agent should generate a migration script and tests, execute them in a test environment, present code and rationale for human approval, then run the deterministic artifact.
Tasklet praised Blaxel for fast-starting sandboxes and Firecrawl for crawling, while retaining its own database and file-system infrastructure because those layers remain core. Credits are deliberately generic enough to buy tokens, search, forthcoming native image generation, and eventually music; excluding user API-call costs, internal token spend is estimated at roughly 5–10% of payroll. On Mythos, Lee’s honest hedge was that the benchmarks and zero-day claims look exciting, but without access it still feels “a little bit…like a marketing stunt.”