Blits.ai vs Sierra (2026)

Agents act in outside systems through three routes. A catalog covers 56 enterprise systems, among them Salesforce, SAP, ServiceNow, Workday, Zendesk, Oracle, Jira, Snowflake and Databricks, and business tools for HubSpot, Freshdesk, Zoho, Notion, Odoo, Dynamics 365 and Microsoft 365 connect through per system OAuth.

Custom functions call REST or SQL sources, or run JavaScript in an isolated sandbox, with versioning and publishing. External MCP servers connect over SSE or HTTP, and their tools appear to the agent. Built in tools add web search, file generation, payment links, lead capture, handover, and call control on voice such as transfer, DTMF and callback scheduling.

Sourceblits.ai/featuresread 2026-10-13

Agents reach the customer's systems of record to carry out end to end transactions, including card and ACH payments, returns, authentication and cancellation saves, and those system interactions are held deterministic and bound by policy. Agent Studio offers ready integrations and an integration framework for proprietary systems, and the Agent SDK builds the rest, so the range of connected systems depends on what each team sets up. Integrations also route a conversation to the customer's care team with its full context.

Sourcesierra.ai/product/agent-sdkread 2026-09-05

A BPMN style canvas lays out each conversation as flows of dialog, action, condition and AI blocks, with subflows, functions and agents referenced from any flow and versions published separately to Test and Live.

The agent block hands a turn to an LLM agent, and the trigger workflow block starts an agentic workflow, an agent loop that calls tools on its own until the task is done. Workflows carry instructions, configuration, a validation check and an on or off switch.

Agentic tasks hold a deferred action of the do X when Y kind, with scheduled rechecks, an expiry and a dry run mode, so a step can wait days for its condition before it runs.

Sourceblits.ai/featuresread 2026-10-13

Agents are put together from skills such as triage, respond and confirm, combined into multistep workflows for a specific use case. In Agent Studio a team defines those journeys step by step or has AI generate them from its existing operating procedures, and Ghostwriter builds or changes an agent from a description of how it should behave. Flexibility is set per workflow, so a team can move between deterministic execution and open ended reasoning on the same agent.

Sourcesierra.ai/product/agent-sdkread 2026-09-05

Fourteen channels carry work to an agent, from the web chat widget, WhatsApp, SMS and phone to Microsoft Teams, Slack, Telegram, Facebook Messenger, email, Alexa and Google Assistant, an in app webview, a digital human and a REST or WebSocket API.

Agentic workflows also start on a schedule, or when an outside system calls a per workflow token endpoint, which queues the run with the request body as its input and allows 30 runs a minute per workflow by default.

Agentic tasks recheck a condition on a schedule before acting, outbound campaigns dial a contact list inside a set calling window, and an incoming email webhook turns mail into conversations.

Sourceblits.ai/api-docsread 2026-10-13

An agent is built once and deployed across chat, phone, email, SMS and messaging, with the same skills and policies applying on each. A single agent working across every channel is the core idea of the platform.

Sourcesierra.ai/product/agent-sdkread 2026-09-05

Retrieval runs over a maintained document library that ingests PDF, Word, PowerPoint, Excel, CSV, JSON, Markdown, images and Outlook messages, crawls websites and their subpages on a recrawl, and keeps version history with revert.

Chunk size and overlap are configurable, storage sits on PostgreSQL pgvector or file based HNSW indexes, and search combines dense vectors with BM25 keyword matching and deduplication. Agents can also query registered PostgreSQL or SQLite databases directly.

Answers cite the document, page and source date, newer content ranks ahead when sources overlap, and retention rules expire outdated material automatically.

Sourceblits.ai/featuresread 2026-10-13

Agents ground answers in the customer's own content through retrieval built into Agent OS, and the Agent Data Platform joins that knowledge with customer context from other systems. In Agent Studio a team connects its knowledge sources and views and edits help center content, FAQs and policies. Sierra spots common themes missing from the knowledge base and drafts new articles from how care representatives resolved edge cases, so the store grows without manual updates.

Sourcesierra.ai/product/agent-sdkread 2026-09-05

A conversation keeps its state across turns and page refreshes, and a client can reconnect to an existing conversation, load or delete one user's history, and push state variables such as a customer name or account number into the session.

Inside a flow, a get history block reads the transcript, and a user data provider turns verified balances, loyalty points and tiers into variables. Data removal and retention controls sit in the GDPR toolkit.

Nothing the agent learns is written to a store of its own for later conversations, so what it knows about a returning customer is the history and the variables the host system supplies.

Sourceblits.ai/api-docsread 2026-10-13

Agents personalize each conversation from agent memory and the customer's profile, and the Agent Data Platform brings together what the agent knows about a specific customer across sessions, channels and systems, so context carries from one conversation to the next. How long that memory lasts, and whether a customer's record of it can be viewed or cleared, is not public.

Sourcesierra.ai/product/insightsread 2026-09-05

Agentic actions above a configurable threshold pause for a person, who approves or rejects them before they run, and a tool execution policy limits which tools an autonomous agent may call at all. Agentic tasks offer a dry run mode, and an anomaly review queue flags unusual conversations for review and can push them into a test suite.

For purchases, authorize then capture holds a payment for later capture, with amount ceilings and velocity limits. Staff can watch live conversations and take over, phone calls included, and blocked topics are enforced deterministically, independent of the model's judgment.

Sourceblits.ai/featuresread 2026-10-13

Sierra wraps its models in supervisory layers that check outputs against scope and policy before the agent responds, to reduce hallucination and prevent abuse. Access to systems of record is deterministic and controlled, and the customer's policies and security procedures are enforced outside the model. Flexibility settings on each workflow limit how far the agent may improvise.

Sourcesierra.ai/product/trust-and-reliabilityread 2026-09-05

Single sign on works with Microsoft Entra ID and Google accounts, with multi factor authentication on admin login, and role based access control offers more than 60 permissions, custom roles per tenant, roles per bot, and separate publish rights for Test and Production. Tenants are isolated down to every record and query, API keys are hashed and validated server side, and PII and card numbers are masked at the gateway before text reaches a model.

Certifications cover ISO 27001:2022, ISO 42001 and SOC 2 Type II, and the trust center also shows GDPR compliance, with policies and controls available on request. Payments run on an isolated service built to PCI DSS v4.0.1 Level 1, with tokenization at the payment provider and encrypted credentials, while PCI DSS certification itself is still in progress.

Sourceblits.ai/featuresread 2026-10-13

Sierra holds ISO/IEC 27001:2022 and ISO/IEC 42001:2023 certificates, a SOC 2 Type II report, a HIPAA attestation for 2026 and PCI DSS Level 1 service provider certification. Its trust center also covers the European and California privacy laws, the CSA STAR registry and FedRAMP High, with controls and policies shared on request.

Personal information shared with an agent is encrypted and masked automatically, filters block topics and keywords a company puts off limits, and cardholder data runs on dedicated PCI infrastructure that never touches Sierra's core platform, models or storage. Customer data is never used to train models or shared with other customers. Who inside a customer's own organization can build, edit or publish an agent, through single sign on, roles or permissions, is not public.

Sourcesierra.ai/product/trust-and-reliabilityread 2026-09-05

Every agent turn can be inspected with token counts and trace detail, including each tool and SQL call with its origin, arguments and latency, and per agent statistics track token use and cost. Agentic workflow runs keep a full audit trail per run with downloadable run data, and the platform records user audit logs and tamper evident event logs. Conversation logs open in a permission gated transcript viewer, live conversations can be watched as they happen, and outbound campaigns log every call attempt.

Sourceblits.ai/featuresread 2026-10-13

Every agent action can be traced to its reasoning, with the knowledge sources and systems it used, and developers inspect API calls and logic traces while Agent Studio shows a live trace of each decision, tool call and response during a build.

Insights reports on CSAT, resolution and other metrics with custom reports, tags and categorizes conversations automatically, flags the ones that need a person's attention and sends alerts on abuse attempts or performance drops to a team's existing tools.

Explorer answers questions about thousands of conversations with themes, data and links to the conversations behind each finding, and delivers a weekly briefing on its own.

Sourcesierra.ai/product/insightsread 2026-09-05

Three deployment routes are offered, a managed service, the customer's own AWS, Azure or Google Cloud account, and an on premises install in the customer's data center, including air gapped sites.

The managed cloud has regional endpoints in the EU, the US, the UAE, Qatar and Australia, each with the same API, and data sent to a regional endpoint stays in that region.

Deployment is Kubernetes native, a self contained Docker Compose stack covers local installs, Google models can be routed to EU or UAE regions through Vertex AI, and transcription can run on GPU servers that Blits.ai operates itself.

Sourceblits.ai/deployment-optionsread 2026-10-13

Sierra hosts the platform as SaaS. Each subprocessor carries its own processing location on the trust center, and customers do not pick a region. There is no VPC, single tenant or on premises option. The FedRAMP High commitment points to separate government infrastructure, though no choice of residency regions is offered.

Sourcesierra.ai/product/trust-and-reliabilityread 2026-09-05

Create with AI turns a written use case into a working bot with agents, flows, guardrails, knowledge suggestions and a go live checklist, and a solution portfolio in the admin console groups curated use cases by sector with their capabilities and channels.

More than 45 rich card types cover flights, hotels, orders, loyalty points, card benefits, balances and charts, each with its own prompt instructions. Each agent is still assembled for the customer, and there is no gallery of prebuilt agents to switch on. The public AI Use Case Library of 282 use cases is research on deployments across the market, not a set of installable agents.

Sourceblits.ai/featuresread 2026-10-13

The Agent SDK ships with a range of AI skills and systems integrations to help customers go live faster, and skills such as triage, respond and confirm work as reusable building blocks. Customers assemble these pieces themselves. There are no named prebuilt agents, templates, library or marketplace.

Sourcesierra.ai/product/agent-sdkread 2026-09-05

An admin chooses the model for each agent from more than 15 providers, including OpenAI, Azure OpenAI, Anthropic, Google Gemini and Gemma, Meta Llama, Mistral, Microsoft Phi, Qwen, DeepSeek, GPT OSS, Hugging Face, Humain, Fanar and OpenRouter, and can move an agent to another provider without rebuilding it. Any OpenAI compatible model connects with the customer's own key.

Regional routing sends traffic to EU or UAE regions, reasoning effort is set per agent, and models above a plan's level stay visible but disabled. Embeddings come from OpenAI, Google, Nomic or OpenRouter, and intent recognition can run on Microsoft CLU, Dialogflow, Watson, Wit.ai or an LLM.

Sourceblits.ai/featuresread 2026-10-13

Agents draw on more than fifteen frontier, open weight and proprietary models chosen per job, and Sierra switches between LLM providers automatically to optimize performance and keep running during a provider outage. Model serving runs through subprocessors that include Anthropic, Google, AWS, Baseten, Fireworks AI, Together AI and Modal. Routing happens inside Sierra, and neither customers nor admins choose the model.

Sourcesierra.ai/blog/constellation-of-modelsread 2026-09-05

A REST API and a WebSocket API drive the platform from outside, authenticated with bot, agent or workflow JWTs sent as bearer tokens to platform.blits.ai or a regional host.

Endpoints open a conversation, send text, speech or files, check whether it is active, erase a user's data and run intent analysis on a message, and the WebSocket adds streaming speech, history loading and state variables. Separate endpoints run an agent with streamed NDJSON output, start an agentic workflow and return metered interaction counts per tenant. Each hosted MCP server has its own URL and exposes an agent's tools to outside MCP clients over Streamable HTTP with an API key.

Sourceblits.ai/api-docsread 2026-10-13

Developers build, customize and deploy agents with the Agent SDK, which comes with documentation, interactive training and a development pipeline. Agent OS adds ready made tooling for building, deploying and managing agents. There is no public API reference or MCP server for the platform itself, so developers work through the SDK.

Sourcesierra.ai/product/agent-sdkread 2026-09-05

Test suites run automated conversations against agents, bots and workflows through the real chat pipeline, and multi turn sets play out as one continuous session with notes per turn for the grader.

Grading uses deterministic rules or an LLM judge with a configurable prompt, a selectable model and knowledge base evidence, and run reports show the pass rate, per turn verdicts, grader reasoning and run history, with regrading. Test cases import from JSON.

Monitors run scheduled health checks against rule based or LLM expectations and alert by email or webhook, and separate checks compare speech to text providers on word error rate and test MCP tool discovery.

Sourceblits.ai/featuresread 2026-10-13

Agent Studio simulates conversations and scores the agent against defined outcomes, and a team builds a regression test suite that runs before each new version of the agent is released.

Voice agents are tested for transcription, background noise and speaker variation before going live, and the Agent SDK checks behavior across a wide range of scenarios in a development pipeline, so definitions are not edited live.

After launch, Insights runs experiments on variations of agent behavior such as handoff rules or levels of determinism, and Explorer compares performance before and after a change Ghostwriter has made.

Sourcesierra.ai/product/agent-studioread 2026-09-05

A web browsing tool sits among the built in agent tools beside web search and Wikipedia, so an agent can open and read web pages while it works. UiPath is one of the 56 connectable systems, and the agentic product connects RPA tools, which lets an agent hand screen based work to a robot the customer already runs. There is no hosted browser or desktop session for the agent to drive itself. Real browser testing, which loads the live chat widget under simulated network and CPU conditions, is quality assurance on Blits.ai's own product.

Sourceblits.ai/featuresread 2026-10-13

Agents act on the customer's systems through API integrations built with the SDK, kept deterministic and bound by policy, and deploy across chat, phone, email, SMS and messaging. Browser control, a virtual desktop and operation of software without a programmatic interface are not part of the platform.

Sourcesierra.ai/product/agent-sdkread 2026-09-05