Runs start without a person
ConfigurableA workflow starts from a schedule, a webhook, an incoming email, a new file, an API call or a manual run. It then works until the goal is reached, a stop condition applies or the time limit runs out.
How Blits.ai agents work
A Blits.ai agent takes a goal, picks the tools it needs, acts in your systems, reads what came back and decides the next step. This page describes that loop, what agents can do, what they remember and where people stay in control.
Each capability is marked as on by default or configurable per assistant, workflow or tenant. Defaults that touch customer data can be switched off where regulation or your own policy requires it.
The agent loop
Workflows are where agents do longer work on their own. Conversations use a faster variant of the same loop.
Plan, act, read the result, decide again
On by defaultIn a workflow the agent works in a loop of up to 25 steps. After every tool call it reads the result and chooses the next action, so the plan changes when a system returns an error, an empty result or something unexpected.
A time limit on every run
On by defaultEach workflow run has a maximum duration, set per workflow, so a run that cannot finish stops instead of running forever.
Agents that delegate
ConfigurableAn agent can pass a sub task to a specialist agent. When an agent judges its own answer insufficient, it hands the question to the agent configured for that domain. A conversation flow can also start a workflow.
Agents in live conversations
On by defaultIn voice and chat, the agent picks the tools it needs in one round and then answers, which keeps response times short. Work that takes longer runs as a workflow, started from the conversation.
What agents can do
Agents act through tools. In workflows, the model decides at run time which tool to call and with which arguments.
Tools chosen at run time
ConfigurableREST calls, SQL queries and code functions are offered to the workflow agent as tools. Agents also call tools on MCP servers and on the built in system connectors.
Custom code in a sandbox
On by defaultCode functions run JavaScript in an isolated sandbox with a 128 MB memory limit and a one second timeout. Code is checked for blocked patterns before it runs.
Files
On by defaultAgents generate files in 13 formats, including PDF, Word and Excel, and send them to the user or pass them to the next step.
Databases
ConfigurableAgents read and filter tenant SQL storage through structured queries, never free form SQL written by the model.
Web
ConfigurableA web browser tool that refuses internal network addresses, plus Google search, Exa search and Wikipedia.
Work with other agents
ConfigurableAgents call external agents over the A2A protocol, and Blits.ai agents can be published for agents on other platforms to call.
State and memory
What an agent knows about the conversation, the customer, earlier runs and goals that span several conversations.
Conversation history
On by defaultEvery message is stored and passed to the agent as context for the next turn.
Long term memory per customer
On by defaultAfter a conversation, the agent keeps lasting facts and preferences, such as preferred language or products held, and recalls the relevant ones in the next conversation. Card data, credentials and special category data are never stored, and every memory can be viewed and deleted per customer.
Switch it off per assistant where regulation or your own policy requires it.
Returning customers
On by defaultWith the setting to continue earlier conversations, a returning customer picks up where they left off, including the state of the conversation.
Switch it off per assistant where regulation or your own policy requires it.
History across conversations
On by defaultA flow step fetches the customer's earlier conversations, optionally summarized by a model, so the agent can use them in the current one.
Switch it off per assistant where regulation or your own policy requires it.
Verified customer data
ConfigurableIn the webview channel, the platform loads verified customer data such as a balance or loyalty points at the start of the conversation.
Workflow runs
On by defaultEvery workflow run is stored with its full trail of tool calls and results, written while the run is in progress.
State across workflow runs
ConfigurableA workflow reads and writes its own state, such as the last processed record or the cases already handled, so each run continues where the previous one stopped.
Goals across conversations
ConfigurableAn agentic task tracks a goal for one customer across conversations, for example following up until an application is complete, with a visible status and an expiry.
Retention and deletion
ConfigurableRetention per knowledge base, 12 months by default for outbound call recordings, and deletion per customer, conversation or assistant.
Where people stay in control
Autonomy where it is safe, fixed steps and approvals where it is not.
Approval per tool
On by defaultTools that move money or change data wait for approval by default, from the customer in the conversation or from an employee, optionally only above an amount. Read only tools run automatically, and every tool can be set differently. Agents can also schedule tasks for later.
Switch it off per assistant where regulation or your own policy requires it.
Fixed flows where wording matters
ConfigurableRegulated steps such as identification, disclosures and consent run as fixed BPMN flows next to the agents, with exact wording and a fixed order.
Handover to people
ConfigurableConversations move to a live agent in Salesforce, Zoho SalesIQ or Freshdesk, with the history attached.
How agents work: FAQ
Want to see it on your own systems?
We run a working session on your use case, with your data and the models you choose.
Plan a call