Agenta

More than 1,000 integrations, connected through Composio, cover mail, chat, calendars, documents, issue trackers, analytics and CRMs such as Slack, Notion, GitHub, HubSpot with read and write access, Google Drive, Linear and Zendesk.

Each app exposes separate actions, so a team can let an agent read an issue but not create one, and connecting an app does not hand over every action it offers.

Any MCP server joins over HTTP, with calls relayed through the API container so the server's credential never enters the agent's sandbox, and an existing workflow can be exposed as a tool with its own inputs and logic.

Sourceagenta.ai/docs/concepts/tools-and-integrationsread 2026-10-18

Agents plan their own steps: the harness loops between the model and the tools until the job is done, compacts long conversations and reads AGENTS.md files in the folders the agent works in.

Delegation works through reference tools, where a parent agent calls another Agenta workflow by name, pinned to a revision or to whatever is deployed in an environment, and the child runs server side with its own connections and secrets while its trace links back to the parent.

Messages can be queued while an agent works, and automations turn a chat job into a recurring unattended run. There is no visual canvas for wiring steps.

Sourceagenta.ai/docs/reference/agents/subagentsread 2026-10-18

Knowledge lives in files. Each agent keeps a persistent folder of brand guides, policies, approved examples and Markdown wiki pages that it opens on demand, so a collection can grow to hundreds of files without enlarging every turn, and a shared AI Drive holds images, PDFs and whole folders that people and agents work from together.

Instructions point the agent at the right files, and subfolders can carry their own AGENTS.md rules. Retrieval is the agent reading files and calling tools such as Google Drive or Notion, with no embedding index, chunking settings or cited search across a large document collection.

Sourceagenta.ai/docs/concepts/files-and-knowledgeread 2026-10-18

Every agent has a default permission, allow reads, allow all or ask, and individual tools can override it, so an agent can read analytics freely but must ask before it publishes or changes a record. Approval cards appear in the app, on mobile, as a click in Slack and as Approve and Deny buttons on WhatsApp, and leaving Always auto approve unchecked limits the approval to that one call. Unattended automation runs pause for a person to fill a form, connect a service or approve an action, then continue once the input arrives, and a user can stop a running turn at any point.

Sourceagenta.ai/docs/changelog/agent-permission-policiesread 2026-10-18

SOC 2 Type II certification, with the report on the Business plan, backs a cloud that encrypts secrets and data at rest with AES 256 and traffic with TLS 1.2 or later. Six roles, Owner, Admin, Developer, Editor, Annotator and Viewer, apply per organization and are enforced in every edition including open source.

SSO runs over OIDC with Okta, Azure AD, Auth0, OneLogin or Google Workspace, can be enforced so password login is off, and pairs with DNS verified domains that route matching users to the identity provider. API keys are scoped to one project, and in Daytona sandboxes provider keys become secrets bound to a single host, so a prompt injection that prints the environment exposes only placeholders.

Sourceagenta.ai/docs/administration/security/overviewread 2026-10-18

Each run is traced as a nested span tree whose IDs come back in headers, response fields and stream events, and tool spans tell a permission denial apart from a client pause, an integration error or a sandbox failure. Streaming runs report input and output tokens and cost, and analytics cover usage and model spend per agent.

Automation runs keep the whole conversation with tool inputs and outputs, generated files, errors, the configuration and the delivery that started the session, and traces from child workflows link back to their parent. Enterprise plans add audit logs of access and changes.

Sourceagenta.ai/docs/reference/agents/traces-and-usageread 2026-10-18

Context carries forward in two ways. Sessions keep their conversation and session files, can be grouped, pinned and reopened, and a WhatsApp customer's thread holds context across their messages. Across sessions, the agent's own folder persists, and it carries learning forward when the team or the agent writes notes, approved examples and lessons into those files and the instructions point back to them. Nothing extracts preferences or facts from conversations on its own, so long term memory depends on someone, person or agent, choosing to write it down.

Sourceagenta.ai/docs/concepts/files-and-knowledgeread 2026-10-18

Teams can run the whole platform themselves under the MIT license, with Docker Compose, Kubernetes production guides or Railway, and choose whether agents execute in the local runner container or in isolated Daytona cloud sandboxes.

Agenta Cloud runs on AWS in an EU or a US region picked at signup, and every prompt, evaluation, trace, secret and account record stays in that region.

Enterprise customers can take a dedicated instance in a custom cloud region or a virtual private cloud, with their own domain, isolation and retention rules, and self hosted agents can run against a Claude or ChatGPT subscription the team already pays for.

Sourceagenta.ai/docs/administration/security/data-regionsread 2026-10-18

The Agent Marketplace offers 29 working agents across engineering, support, sales, knowledge and operations, among them a PR reviewer, changelog writer, issue triage, CI failure triage, incident responder, support reply drafter, lead qualifier, CRM updater, meeting follow up and weekly report.

Using one creates an agent with the template's instructions, skills and files, then asks for the accounts it needs before the first task. Teams can save their own agent as a template zip that leaves secrets out and keeps automations off, or have the agent open a pull request to share it in the marketplace, and skills published once install into any agent with versioned updates.

Sourceagenta.ai/marketplaceread 2026-10-18

Automations start an agent on a five field cron schedule or on an event from a connected app, such as a new GitHub issue or a new row in a sheet, with JSONPath selectors pulling event fields into the starting message.

Agents publish to Slack, where they answer mentions in any channel or thread, to Telegram by QR code or a team's own bot token, and to a WhatsApp business number, where each customer's phone number is its own conversation and replies outside Meta's 24 hour window wait until the customer writes again. A REST invoke endpoint, webhooks and a mobile app round out the ways work reaches an agent.

Sourceagenta.ai/docs/reference/agents/automate-agent-runsread 2026-10-18

Harness, model and credential are three separate choices. Claude Code runs Anthropic models directly, through AWS Bedrock or Google Vertex AI, or through a company gateway; Pi takes OpenAI, Anthropic, Google Gemini, Mistral, Groq, MiniMax, Together AI, OpenRouter, xAI and any OpenAI compatible endpoint; and Codex runs five OpenAI models.

Each agent picks its own model, Agenta's included models work without a key, Ollama covers self hosted models, and a team's own API key or ChatGPT subscription removes model charges, so switching any piece is a configuration change that leaves instructions, skills and tools in place.

Sourceagenta.ai/docs/concepts/harnesses-and-modelsread 2026-10-18

A public REST API, described in an OpenAPI 3.1 spec and served from US, EU or self hosted base URLs, invokes agents, starts and continues sessions, reads per turn events, uploads files, manages triggers and schedules and queries traces, authenticated with project scoped API keys. The Python SDK installs with pip, and agent output also streams in Vercel AI SDK format. Agents extend through MCP servers and skills, and because the code is MIT licensed, teams can customize the agent runtime image, add sandbox providers or contribute changes upstream.

Sourceagenta.ai/llms.txtread 2026-10-18

Evaluation came from Agenta's first product and now covers agents: versioned test sets, evaluators that range from LLM as a judge with custom output schemas to regex, semantic similarity, RAG metrics, webhooks and custom code, and persisted evaluation runs over a test set.

The playground attaches evaluators inline and compares variants side by side, annotation queues let reviewers label traces and turn failures into test cases, and the SDK runs evaluations from code. Agent configuration history compares any two versions of instructions and tools and restores an earlier one, and a new automation can be test run on the spot.

Sourceagenta.ai/docs/changelog/playground-evaluation-workbenchread 2026-10-18

Each agent gets its own computer, a sandbox with 2 vCPUs and 4 GiB of memory, and the sandbox image ships Playwright with a headless Chromium alongside Python and Node.js, so an agent can write and run scripts that open sites, read pages and step through them. A browser panel sits next to the files in the agent's workspace for research work. There is no dedicated browser tool with a live view, handover to a person or saved logins for signed in sites, and no control of a desktop beyond the sandbox's own shell.

Sourceagenta.ai/pricingread 2026-10-18