Definition: A Taskade autonomous agent is an AI Agent v2 instance that runs inside your workspace, has persistent memory across sessions, can call any of 22+ built-in tools (web search, code execution, file ops, integrations), and triggers automations on events. Unlike open-source frameworks that ship a runtime and ask you to build everything else, Taskade ships the runtime and memory, identity, observability, and the workspace itself.
What Makes Them Autonomous
A Taskade agent isn't just a prompt-and-respond chatbot. Each agent has:
- Soul (instructions + persona) — the brief that defines what the agent is for
- Tools — slash commands, custom commands, public tools (
/web,/code, integrations), workspace tools (read project, update task, send Slack) - Memory — workspace-scoped knowledge: documents, projects, prior conversations
- Triggers — events that wake the agent up (a webhook, a schedule, an inbox arrival, a button click)
- Audit trail — every action logged with attribution and timestamp
- Permissions — 7-tier RBAC controls who can see, edit, or invoke
Put together, this is a software entity that takes initiative, remembers, and is accountable. That's the autonomous part.
Taskade Agents vs Open-Source Claws
There's a related concept — AI claws — coined by Andrej Karpathy in March 2026. Claws are persistent autonomous agents you self-host (OpenClaw, NanoClaw). They give you maximum control and a maximum security burden. Taskade agents give you a managed environment: less control over the runtime, far less security/scaling burden, and a workspace built for collaboration. See /wiki/autonomous/ai-claws for the comparison.
Capabilities Shipped Through v6.164
Genesis has shipped a lot of agent capabilities in the last year. Highlights from v6.140 → v6.164:
| Version | Capability | Who It's For |
|---|---|---|
| v6.150 | Ask-Questions tool — agent can pause and ask clarifying questions mid-task | Anyone running long-running agent workflows |
| v6.150 | Auto Top-Up — credits auto-purchase when usage crosses a threshold | Teams running agents 24/7 |
| v6.151 | App Users (beta) — first-class end-user management for published Genesis apps | Teams shipping agent-powered apps to non-Taskade users |
| v6.153 | Memory Graph (/memory) — visualize the Workspace DNA knowledge graph |
Anyone debugging agent context |
| v6.153–6.155 | Bash Sandbox + VFS-backed Bash Sandbox in Space Agents | Agents that run real code |
| v6.154 | Custom Bash Commands — user-defined shell scripts as agent tools | Teams with bespoke ops scripts |
| v6.160 | 15+ frontier models via gateway routing (Qwen 3.6 Max, Kimi K2.6, DeepSeek V4 — open-weight) | Teams that want broader model choice |
| v6.161 | utils.jsonExtract (JSONPath) and setProjectTitle action |
Automation authors |
| v6.162 | ask_ai_structured — typed JSON output via user JSON Schema |
Anyone wiring agent output into downstream systems |
| v6.162 | website.summarize — one-click URL-to-summary action |
Research and lead-gen workflows |
| v6.163 | Bulk-Delete UI for project listings + agent reference libraries | Teams managing large agent libraries |
| v6.163 | Enterprise BYOK — bring your own OpenAI / Anthropic key | Enterprise customers with strict data policy |
Multi-Agent Teams
Single agent ≠ best agent. Many problems are better solved by a team of agents — a researcher, a writer, an editor — each with a narrow role and clear hand-off rules. Taskade ships multi-agent teams as a first-class primitive: agents can call each other, share Project memory, and cooperate on a single task. See also the quantum-style multi-agent interference for what happens when teams run in parallel instead of sequentially.
Triggers (How Agents Wake Up)
| Trigger | Example |
|---|---|
| Schedule | "Every Monday at 9am, run the weekly digest agent" |
| Webhook | "When Stripe sends a payment event, run the AR agent" |
| Inbox arrival | "When a new Gmail with subject Lead: lands, run the SDR agent" |
| Calendar | "When a Calendly booking is created, run the prep agent" |
| Button click | "When a user clicks Suggest reply, run the support agent" |
| Time on field | "If this lead has been in Prospecting for 5 days, nudge the SDR agent" |
| @-mention | "When someone @-mentions me in a project, react" |
Every trigger maps to an Automation; every Automation can hand off to one or more agents. That's the closure of the autonomous loop.
Cost Model
Free tier includes a one-time AI credit allocation; paid plans add monthly credits, and credit packs use volume-tiered pricing. See the canonical breakdown on Pricing and Credits — those pages stay in sync as plans evolve, so we don't pin specific numbers here.
For 24/7 agent workloads, Auto Top-Up (v6.150+) prevents agents from stalling mid-loop when credits run out.