Product
Agents
Agents built from templates or from a goal you describe. They run on a schedule, on an event or on demand, and show every step they took. Their writes go through the policy engine before they happen.
Agents
- Agent and playbook template library with categories and role filters (Founder, Sales, Marketing, RevOps, CS), curated and community sections, one-click share and reuse.
- Standard template anatomygoal, numbered steps, who it's for, inputs, outputs, integrations with permission scopes, approvals, when it runs, what it replaces, estimated savings and FAQ.
- Describe-a-goal agent builder that plans and runs multi-step work with visible step status.
- Trigger typesschedule (cron), event (CRM change, deal closed-won, meeting recorded, new email), inbound (form, demo request, identified visitor), webhook, manual or on request.
- Overnight scheduled skills that prepare leads, drafts and content into a morning queue.
- Durable executionagents do a bounded unit of work per run while the platform owns long-lived state (resumable runs, retries, follow-up cadence, full run history, replay, alerts).
- Run trace where every step is a tool call tagged read, reason or write, with a note saying what ran.
- Per-integration permission scopes labeled Reads / Writes new / Modifies existing, with optional assisted access, chosen before deploy.
- Tool-level gatesrequire approval before a specific tool runs, disable tools, restrict tools to admins.
- Approval queue for agent-generated actions with pending and completed counts and one-click approve, edit or reject.
- Per-contact whole-sequence approval with inline per-step edits, available and kept in sync across the app (J/K keyboard review), a chat card and Slack buttons with an edit modal.
- Approval expiry (default 10 days, configurable 1–90) with explicit status codes.
- Autonomy levels from approve-every-message through per-contact approval and explicit per-sequence auto-send to full autopilot within guardrails.
- Staged writes. Cascading actions summarize first and wait for confirmation. Merges, deletions, schema changes, owner changes, lifecycle updates and paid enrichment stay blocked until approved. External API/MCP callers cannot bypass owner approval.
- Uncertainty escalationthe agent flags low confidence and waits, irreversible actions always confirm, and a queue holds tasks that need a person's decision.
- Guardrails enforced in the send layer regardless of model output (suppression, do-not-contact, existing-customer and open-deal checks, deterministic controls).
- Learning from human editsan edited first message is saved as a marked example; Slack corrections retrain; skipped accounts and rewritten emails captured as team context.
- Outcome-driven evalseach playbook version measured on replies, meetings, attendance, closed-won and sign-ups against version n−1; a regression benchmark gates every harness change.
- Pre-launch simulations, auto-QA, live A/B of agent versions with winner promotion, conversation assessments and response scorecards.
- Versioned, orchestrated agent deployment with version control.
- Agents as managed staffname, title, job description, org-chart position, a human manager, skill slots and run frequency.
- Skills as portable Markdown files (SKILL.md) and installable skill packages or public skill libraries.
- Deterministic code for predictable steps with LLM judgment reserved for hard stepsdeterministic code sandbox; isolated per-agent cloud environments; agent writes and runs code.
- Model routing by task cost and quality, a model your organization picks, and your own key if you bring one. Anthropic's models today.
- Spend governancedaily budgets, cost confirmation before rewrites or data purchases, credit-balance stop that pauses rather than overcharges, live usage by category.
- Agent lifecycle states (Draft, Starting, Working, Waiting on you, Done; Live, Ready) and run counters on agent cards.
- Persistent agent conversations that can be handed to a teammate.
- Project workspaces combining chat, context, audiences, files, people and canvases, shared with the team.
- Business-logic knowledge base (positioning, proof points, voice, ICP, playbooks) that agents draw on.
- Per-account persistent agents with memory, decisions and strategy exposed through an API.
- Multi-agent orchestrationspecialized agents for finding, writing and designing, a router that hands them work, and separate builder and worker agents.
- Self-healing workers, and proposals for new angles or prospect pools when an agent's run quality dips.
- Workflow builder with an agent step, rules engine, recipes and outbound HTTP/webhook blocks.
- Explainability on every outputagents show their work, give the reason on each row, and cite the evidence behind each recommendation.
- Recovery trail and dry-run for agent writes (verify every write; `--dry-run` and `--force` on destructive calls).
- Auto-clear: an action or flag resolves itself once data or a reply shows the issue is fixed.
Other areas
Try it with your own data
Free for up to 3 seats, no card.