Tasklet gives you a text box. "Describe your goal," it says. No canvas. No drag-and-drop nodes. No if-then-else branches to wire up. You type what you want done, something like "review all my subscriptions across email receipts and bank statements, identify unused ones, and cancel them using the browser," and the agent figures out how to do it.
Most products in this series still give you some kind of builder. Lindy gives you a visual agent builder with trigger nodes and action steps. Claude Code routines give you a prompt-plus-triggers model anchored to code repos. Zapier and Make give you the workflow canvas that defined the automation category for a decade. Tasklet skips all of it. The founding team, Andrew Lee and Jonny Dimond, previously built Firebase and then Shortwave. They call this philosophy "always bet on the models." Let the AI plan the whole thing.
Tasklet launched in October 2025, crossed $5 million in annual recurring revenue by early 2026, and raised $20 million from Union Square Ventures and Lightspeed at a $175 million valuation. USV's blog post about the investment includes a detail that says more than the numbers: the firm signed an enterprise contract and rebuilt their own operating system on Tasklet before the round even closed.
Human-defined steps vs. AI-planned execution.↗
The inversion
Traditional automation platforms work on a model where humans define the workflow and AI fills specific boxes. You build the pipeline: "when this happens in Gmail, extract this field, look it up in Salesforce, format the result, post it to Slack." The logic is yours. The execution is deterministic. If the pipeline breaks, you know exactly which step failed and why.
Tasklet inverts this. You describe the outcome, and the AI decides what steps to take, which integrations to call, and how to handle edge cases along the way. Under the hood, Tasklet runs a two-tier agent architecture: a persistent high-level agent that maintains instructions, memory, and user context across sessions, and sub-agents that spin up for individual task executions with scoped permissions. The high-level agent remembers what worked last time. The sub-agent handles this particular run.
The system routes tasks across multiple frontier models: Claude from Anthropic, GPT from OpenAI, and Gemini from Google. Context management uses SQL databases rather than stuffing everything into system prompts, with a file-system-based approach that compresses older history into summaries while keeping recent interactions at full fidelity. An agent that handles thousands of triggers per year can still maintain coherent context without regenerating it from scratch on each run.
The trade-off is legibility. When a Zapier workflow fails at step 4, you can see step 4 and fix it. When a Tasklet agent takes an unexpected path, you're reading a conversation transcript to understand what happened. For teams that need deterministic behavior, that's a real cost. For teams that are tired of maintaining brittle workflows that break every time an API changes, it's the whole point.
More trigger types than any other product in this series.↗
The trigger ecosystem
Tasklet's trigger menu includes: schedules (daily, weekly, custom cron), webhooks, email (Gmail with filters), Slack messages, Google Calendar events, Google Drive uploads, Outlook events, Telegram messages, YouTube uploads, GitHub events, HubSpot changes, RSS feeds, Apple Shortcuts, and text messages over iMessage and RCS.
Fourteen distinct trigger types. For comparison, Claude Code routines ship three (schedule, GitHub events, API endpoint), Lindy has broad coverage but many triggers poll underneath, and Aeon has one.
Mapped against the three primitives:
Clock. Schedule triggers handle the basics. Daily briefings, weekly reports, custom cron intervals. Nothing novel here, but the execution is clean. Agents run in cloud sandboxes on Tasklet's infrastructure with 2 vCPUs and persistent filesystems, so there's no laptop-open requirement.
Listener. This is where Tasklet's breadth stands out. Gmail triggers use Google's push notifications. Slack triggers fire on new messages. RSS feeds push on new items. GitHub events arrive via webhooks. YouTube channel triggers run on new videos. The question with any product that claims this many event-driven integrations is how many are genuinely push-based vs. polling underneath. Tasklet's documentation doesn't always make the distinction explicit, but the latency on email and Slack triggers suggests real push delivery for the core integrations.
Inbox. Two-way communication through iMessage, RCS text messages, email replies, and Apple Shortcuts. You can text your agent while walking to a meeting and get a response. The Apple Shortcuts integration means you can trigger agents from Siri, CarPlay, or location-based automations on your phone. The inbox here isn't just "the agent can reach you." You can reach the agent from wherever you already are.
Most products in this series are strong on one primitive and thin on the others. Tasklet covers more surface area across all three than anything else we've evaluated.
Four fallback tiers: from OAuth down to browser automation.↗
The integration strategy
Tasklet claims 3,000+ integrations, a number that deserves unpacking. The system uses a four-tier strategy:
Tier 1: Pre-built OAuth integrations. These are the named integrations, around 15 with first-class OAuth flows: Gmail, Slack, Salesforce, HubSpot, Stripe, Airtable, Linear, and others. Credentials are encrypted and never exposed to Tasklet's own systems.
Tier 2: HTTP API fallback. For services without pre-built integrations, the agent can call any HTTP API directly. Lee has described how the models generate their own tool descriptions from web-scraped API documentation. You tell the agent "look up this company in our CRM," and it figures out which endpoints to call. The reliability here scales with model capability, which is the founding thesis put to work.
Tier 3: MCP servers. Tasklet supports connecting to any Model Context Protocol server, including private ones your company built internally. Lee's public take on MCP is notably measured: he argues that direct API connections outperform MCP in most real-world usage because modern models can reason about HTTP APIs independently. MCP's remaining value, in his view, is in OAuth authentication flows and edge cases where a standardized abstraction layer genuinely helps.
Tier 4: Browser automation. When no API exists, the agent opens a browser. It can fill forms, navigate authenticated sessions, extract data from web pages, and complete multi-step web workflows. Cookies persist across sessions. This is the fallback of last resort, but it means the agent can interact with enterprise tools that have no API surface at all.
The tiered approach is a pragmatic answer to the integration cost problem this series has covered extensively. Instead of building and maintaining thousands of bespoke connectors (which cost Lindy over a million dollars for 250 integrations before the Pipedream partnership), Tasklet uses model intelligence as the integration layer and invests in first-class connectors only where OAuth requires it.
All three primitives present, nine months in.↗
Running the scorecard
Clock: Yes. Clean schedule triggers with daily, weekly, and custom cron support. Agents run in cloud sandboxes, so the clock doesn't depend on your machine being awake. The persistent filesystem means a scheduled agent can pick up where it left off. Compared to Aeon's GitHub Actions workaround or Hermes's self-hosted gateway, the managed compute is a genuine convenience.
Listener: Yes, with the broadest coverage in the series. Fourteen trigger types spanning email, chat, calendar, file storage, code, video, and RSS. The named triggers appear to use genuine event-driven delivery for integrations where push is available (Gmail, Slack, GitHub, webhooks). For the long tail of 3,000 integrations where events arrive through the HTTP API fallback or browser automation, the picture is less clear. An agent that monitors a website by scraping it on a schedule is polling, even if the cron trigger that kicks it off is clean. The trigger breadth is real, but not every integration delivers events in real time.
Inbox: Yes, and notably bidirectional. iMessage, RCS, email, and Apple Shortcuts give the user multiple ways to reach the agent on their own terms. The text messaging integration is strong enough that USV described delegating to Tasklet agents via text as part of their daily workflow. Combined with the "Instant Apps" feature, which lets agents generate lightweight interactive web apps with live data from connected services, the delivery surface is broader than any product in this series except possibly Lindy's iMessage channel.
The clock is solid. The listener is the most comprehensive in this series but varies in push-vs-poll depth depending on the integration tier. The inbox is genuinely bidirectional. For a product that launched nine months ago, the primitive coverage is notable.
Who should pay attention
Tasklet works best for teams that are already frustrated with maintaining Zapier workflows and are willing to trade deterministic execution for flexibility. The natural language interface means non-technical operators can build agents that would previously require a developer to wire up. If your use case is "handle recurring business operations across multiple tools," and you're comfortable with an AI that plans its own steps, Tasklet covers more ground than anything else in this space right now.
It works less well for teams that need audit trails today (SOC 2 is still in progress), that require deterministic behavior for compliance workflows, or that need predictable per-task pricing. If you need to explain to a regulator exactly what steps your automation took and why, the model-plans-the-workflow approach is a harder sell than a visual canvas you can screenshot.
Lee has been transparent about the competitive picture: the biggest threat isn't other agent platforms, it's the AI labs themselves. Claude Max, ChatGPT Pro, and Gemini Advanced all offer increasingly capable agents with deep integrations to their own ecosystems. Tasklet's advantage is breadth: fourteen trigger types, 3,000 integrations, the four-tier fallback strategy, and the persistent agent memory that accumulates institutional knowledge over time. None of the lab-built agents match that surface area yet. Whether that moat holds as the labs expand their own agent infrastructure is the open question for every horizontal platform in this space.
✦ Newsletter
Liked this essay?
Get the next one in your inbox. One email per essay, no spam.
Posted June 16, 2026 · AgentWorkforce
Issues, PRs, and arguments welcome on GitHub. Or email [email protected].