Agents that finish the job.
100+ tools, multi-agent delegation, self-improving memory, and human approval on every consequential write — orchestrated by a graph engine with hard cost caps and resumable runs.
Execution Engine
Every run takes the same governed path.
A state machine with conditional edges: agents plan before acting, call tools in parallel, and retry retrieval when confidence is low — with the budget gate in front of the first token.
Tools
100+ ways to touch your stack.
Reads run freely. Writes carry risk tiers and wait for approval. Extend with custom webhooks and MCP servers.
search_documents, search_web, recall_memory
NL-to-SQL over PostgreSQL, MySQL, ClickHouse, SQL Server
Gmail, Calendar, Drive, Docs, Sheets — sends from your account
OneDrive, Outlook, Teams, SharePoint, Planner
Zendesk tickets + ServiceNow incidents, changes, CMDB
Salesforce SOQL, HubSpot records & notes, Zoho
Datadog logs & monitors, Sentry issues, PagerDuty
GA4 reports, Mixpanel funnels, PostHog HogQL
Sandboxed Python — no network, secret-scrubbed, approval-gated
Bases, pages, databases, records, comments
User-defined endpoints with configurable auth
Model Context Protocol tools, resources, and prompts
Delegation
A coworker with coworkers.
Agents hand subtasks to specialist agents on the same team — a research agent delegates SQL questions to a database analyst, support triage asks the code specialist about a stack trace. The parent stays in control and integrates the results.
- discover_agents and delegate_to_agent — specialists resolved from the team's delegatable automations
- Specialist budgets clamp to the parent's remaining iterations and cost; spend rolls up to the parent run
- A specialist's approval-gated write pauses the parent — one approval, then the run resumes
- Read-only parents can never gain write tools through a specialist
- Depth-limited to two levels, with a per-run delegation cap
Memory
Self-improving agent memory.
Three layers work together so agents get smarter the more you use them — without operator intervention.
Layer 1
Persistent memory
Facts, preferences, insights, and procedures stored across sessions with importance scoring. Agents recall mid-execution via the recall_memory tool.
Layer 2
Observational memory
After every long turn, a lightweight model distills findings, tool usage, and pending items into a compressed summary — 80%+ token compression on 50-message conversations, cacheable so prompt caches hit.
Layer 3
Procedural memory
Successful multi-tool sessions distill into reusable playbooks. The agent recognizes similar jobs and reuses the steps that worked — admins curate, publish, and version the library.
Team-level or user-level memory isolation
Configurable TTL with automatic cleanup
Periodic similarity-based dedup of overlapping memories
Rate limiting, deduplication, type validation
Budgets
Predictable costs. Resumable runs.
Every run is capped before the first token is spent. If a budget exhausts mid-research, the agent delivers a partial answer and a Continue button — nothing is lost, nothing re-runs from scratch.
Dynamic budget tiers
The orchestrator classifies each query and applies the right tier — upward only, so operator floors are never shrunk.
Pre-flight cost ceiling
Worst-case cost is estimated before the first LLM call; runs over the per-run cap are rejected immediately. A live meter shows spend as the run progresses, plus optional rolling 24h team caps.
Pause & resume
Exhausted budgets checkpoint the full run state for 7 days. Continue extends the budget and picks up exactly where it left off; automations auto-resume on the next tick.
| Tier | When it fires | Iterations | Tokens | Wall-clock |
|---|---|---|---|---|
| Factoid | Simple lookup questions | 10 | 50k | 2min |
| Procedural | How-to / step-by-step | 15 | 80k | 3min |
| Exploratory | Open-ended research | 25 | 150k | 5min |
| Comparative | N-way comparisons | 30 | 180k | 6min |
| Automation | Scheduled deep research | 50 | 300k | 10min |
Reliability
Built for the long tail of real work.
Every agent ships with the resilience features on by default — no flag-flipping, no extra wiring.
Approval-gated writes
Every tool carries a risk tier (T0–T5). Consequential writes pause the run and route to the right human — even when a delegated specialist is the one writing.
Auto-recovery
Transient tool failures (timeout, rate limit, 5xx) retry once before escalating to the LLM. Recovery retries don't count against the tool-call budget.
Context compaction
Long conversations are summarized in place before they crowd out working memory — token-aware, triggering at 50% of the context window.
Verification contracts
Automations can require minimum confidence, minimum sources, or a non-empty answer — agents self-evaluate before completing.
Canvas
Work products, not just answers.
Agents draft versioned artifacts — reports, code, plans, JSON — in a side panel with diff view, version history, and direct editing.
- Markdown, Python, Go, TypeScript, JSON
- Version history with diff comparison
- Edit artifacts directly; the agent revises from there
Extensibility
Teach it your internal tools.
Wrap any internal endpoint as a tool, or plug in MCP servers — new tools inherit the same risk-tier and approval policies as the built-ins.
- Custom webhook tools with configurable auth and timeouts
- MCP servers over SSE, streamable HTTP, or stdio (self-hosted)
- Sandboxed Python for analysis the tools don't cover
Automations
The agent, on the clock.
An automation is an agent with a job description and a trigger. The agent can even draft and manage its own — gated by approval, per team.
Triggers
Cron schedules, inbound email, webhooks from Jira / Zendesk / GitHub / Salesforce / HubSpot, and meeting events.
Verification
Acceptance criteria (min confidence, min sources), stale-state checks, and bounded auto-resume on paused runs.
Delivery
Results land via webhook, Slack channel, or email — with attribution footnotes on external writes.
Templates
5 ways to start.
Pre-built blueprints with system prompts, tool selections, and knowledge base connections. Clone and make them yours.
Support Triage Agent
Classifies tickets, searches the knowledge base, drafts responses
Sales Briefing Agent
Compiles prospect research from CRM, email, and documents
IT Helpdesk Agent
Resolves common issues, escalates complex problems
Compliance Reviewer
Checks documents against policy, flags violations
Onboarding Assistant
Guides new hires through setup, answers policy questions
Get Started
Build your
first agent.
Start from a template or from scratch — see it plan, act, and ask for approval in the live demo.