Agno

The Agno SDK ships more than 100 toolkits, alongside MCP tools and custom tools that agents call to act in outside systems.

SourceAgno, agno.com homepage and docs.agno.com/tools/mcpread 2026-09-21

The SDK gives three patterns, agents, teams that delegate across member agents, and workflows that run ordered steps with loops, parallel branches and approval gates, so deterministic steps and agent steps combine in one run.

SourceAgno, agno.com homepage and docs.agno.com/workflowsread 2026-09-21

Knowledge loads files, URLs or raw text into a knowledge base backed by a vector store, and the agent searches it to ground its response; a published pages mode keeps Markdown and vectors in sync with a documentation site and shows the page revision behind each result.

SourceAgno, docs.agno.com/knowledge/overviewread 2026-09-21

Human-in-the-loop covers agents, teams and workflows: user confirmation requires explicit approval before a tool executes, user input and dynamic user input pause for information, and external tool execution hands a call outside the agent's control, with approvals managed from the AgentOS control plane.

SourceAgno, docs.agno.com/hitl/overview and agno.com homepageread 2026-09-21

AgentOS documents a deep access surface: JWT authentication across REST, MCP, A2A and WebSocket requests, roles as named bundles of scopes, service accounts and per-user data isolation, with basic RBAC on Pro, a SAML SSO add-on, and per-resource RBAC and audit logs on Enterprise. Compliance evidence is thinner: the site shows a SOC 2 badge with no report, auditor or date, and the attestation details sit in a trust center at trust.agno.com.

SourceAgno, docs.agno.com/agent-os/security/overview and agno.com/pricingread 2026-09-21

Tracing records execution spans in the customer's own database, covering run input and output, model calls, tool calls with timing, results and errors, team delegation and workflow steps, viewable in the control plane or queryable through the AgentOS API; retention is unlimited on every plan and Enterprise adds audit logs.

SourceAgno, docs.agno.com/agent-os/tracing/overview and agno.com/pricingread 2026-09-21

Memory stores facts per user_id in the configured database and loads them into later runs across conversations, with a memory mode controlling when records are created, updated or deleted, alongside sessions scoped to a single conversation.

SourceAgno, docs.agno.com/memory/overviewread 2026-09-21

The open source runtime runs in the customer's own cloud, with documented deployment templates for Docker, AWS, Azure, GCP, Helm, Modal, Railway and Render; the pricing page states data never leaves the customer's system, and Enterprise adds a self-hosted control plane.

SourceAgno, agno.com/pricing and docs.agno.com/deploy/templatesread 2026-09-21

Agno publishes complete agent applications as templates, each with its own repository: Scout, a company intelligence agent across web, Slack, Drive, wiki and CRM sources; Dash, a self-learning data agent; and Coda, a code companion for pull request review and issue triage in Slack. Each one stands alone and does its own job.

SourceAgno, docs.agno.com/deploy/templatesread 2026-09-21

The Scheduler creates cron jobs that execute AgentOS endpoints and tracks their run history, so an agent, team or workflow starts without a person initiating it, and AgentOS interfaces serve agents in Slack, WhatsApp and other chat apps.

SourceAgno, docs.agno.com/scheduler/overview and agno.com homepageread 2026-09-21

The developer sets each agent's model by choosing a provider model class and model id, from a documented catalog spanning OpenAI, Anthropic, Google, open-weight hosts and gateways, and the homepage shows agents on different models side by side, so model choice sits with the customer.

SourceAgno, docs.agno.com/models/overview and agno.com homepageread 2026-09-21

Agents built with the Python SDK are served by AgentOS over a documented REST API, as an MCP server, over A2A and through chat interfaces, with a published API reference and authentication on each route, so outside callers drive the platform.

SourceAgno, docs.agno.com/agent-os/security/overview and agno.com/pricingread 2026-09-21

Evals turn expected behavior into repeatable checks that run in development and CI: an accuracy eval runs the agent and has an evaluator model compare its response with the expected output, alongside agent as judge, reliability and performance evals.

SourceAgno, docs.agno.com/evals/overviewread 2026-09-21

Interactive browsing in Agno runs through third party engines such as Browserbase, Stagehand and AgentQL, each bought separately. The product has no browser, desktop or computer control of its own, and its Web Browser toolkit opens URLs in a browser without operating the page.

SourceAgno, docs.agno.com/tools/toolkits/others/web-browser and /tools/toolkits/web-scrape/browserbaseread 2026-09-21