How Databricks and Replit build AI in 2026, and how we plug into them

Most clients arrive with one of two constraints. Either the data is the hard part — it is governed, regulated, large, and already in a lakehouse — or delivery speed is the hard part, and the team needs a working application in front of users this month. Databricks and Replit sit on opposite ends of that split, and both changed substantially in 2026.

The scale is worth stating plainly. Databricks reported a $7 billion revenue run-rate with over 80% year-over-year growth on 13 August 2026, alongside a $5 billion raise at a $190 billion valuation. Replit raised $400 million at a $9 billion valuation on 11 March 2026, reporting over 50 million users and users at 85% of the Fortune 500; its September 2025 round of $250 million at $3 billion followed ARR growth from $2.8 million to $150 million in under a year, with over 500,000 professional users reported at that point.

The two are not alternatives in every case. They are increasingly connected, and a growing share of our work is making one hand off cleanly to the other.

Start with the naming, because it will save you an afternoon. The platform page now brands the product as the 'Databricks Data + AI Platform' with the tagline 'Accelerate the agentic enterprise', and the 'Mosaic AI' umbrella no longer appears in that positioning — though it survives on some pricing and docs pages, including the public Foundation Model Serving pricing page. Mosaic AI Vector Search has been renamed Databricks AI Search. Material written in 2024 or 2025 will use product names that no longer match the console.

Agent Bricks is the centre of gravity. It launched in beta on 11 June 2025 as an auto-optimising agent builder that synthesises evaluation data and custom LLM judges, then searches over prompt engineering, fine-tuning, reward models and TAO. By April 2026 Databricks had recast it as a governed enterprise agent platform, and at the Data + AI Summit on 16 June 2026 reported 100,000+ agents built, processing over one quadrillion tokens per year. That summit ran 15-18 June 2026 at Moscone Center with 30,000+ in-person attendees and 800+ breakout sessions.

The model layer is deliberately multi-vendor. Foundation Model APIs now serve Anthropic Claude Opus 5, Sonnet 5, Fable 5.1 and Haiku 4.5 alongside OpenAI GPT-6 Astra and the GPT-5.x family, Google Gemini 3.x, Llama 4 Maverick, Qwen 3.5, xAI Grok 4.6, Moonshot Kimi K3, DeepSeek V4 and GLM 5.3. That catalogue is the product of deals: a five-year Anthropic partnership announced 26 March 2025, a $100 million OpenAI partnership announced 25 September 2025, and a SpaceX partnership bringing Grok natively to Databricks announced at the 2026 summit. Model launches now land fast — Claude Fable 5 became available on Databricks on 9 June 2026, governed through Unity AI Gateway.

The business-user surface is Genie. Databricks One reached workspace-level general availability in February 2026, giving non-technical users a single pane over dashboards, Genie and custom apps without exposing compute, notebooks or pipelines. On 16 June 2026 Databricks announced Genie One, an agentic coworker with schedules, alerts and native chat-tool surfaces; Genie Agents, domain agents that reason over unstructured data; and Genie Ontology, an automatic business knowledge graph built from tables, queries, dashboards and pipelines. Databricks has not published an explicit general-availability designation for the three; its docs list Genie One and Genie Agents as free through 31 January 2027.

Read the accuracy claims for what they are. Databricks' internal benchmark on 28 real-world enterprise analysis questions reported Genie at 84.5% first-attempt accuracy versus 52.4% for the strongest competing coding agent, at twice the speed. Separately, Databricks claims business-context grounding gives Agent Bricks agents 70% higher accuracy than standard RAG and a 30% improvement on multi-step workflows. These are vendor figures on vendor-selected questions. They are a reason to run your own evaluation, not a substitute for one.

Governance is the part of the platform Databricks is betting enterprises will pay for. At the 2026 summit, Unity AI Gateway took unified AI spend visibility, cost attribution, hard spend caps and smart routing to general availability. Contextual Service Policies — allow, deny or require-approval rules on actions like file modification, code pushes and enterprise system access — are Beta, as are the LLM-based guardrails covering PII exposure, unsafe content, prompt injection, data exfiltration and hallucinations. On the semantic side, Unity Catalog Domains reached Public Preview and Metrics gained materialisation and multi-fact relationships in Public Preview, while Business Glossary was still listed as coming soon and Governance Hub as Private Preview.

One data point is worth carrying into architecture decisions: Databricks says 63% of its customers route across two or more model families. Single-vendor model assumptions are already the minority case.

Replit's arc is legible in four agent releases. Replit Agent shipped 16 September 2024 in early access for Core subscribers, explicitly labelled alpha. Agent v2 followed on 25 February 2025, powered by Claude 3.7 Sonnet. Agent 3 launched 10 September 2025, running autonomously for up to 200 minutes and testing apps in a real browser. Agent 4 arrived 11 March 2026 with an infinite Design Canvas, parallel task execution in isolated exact copies of the project, and plan-while-building; it replaced Agent 3's fork-and-merge collaboration with a single shared project and agent-assisted merging.

The engineering underneath is more interesting than the feature list. Agent 3's self-testing is not a computer-use model: the agent writes and runs Playwright code inside a sandboxed notebook with persistent state, at a median of roughly $0.20 per testing session versus roughly $0.50 for a computer-use model filling a simple five-field form. On evaluation, Replit built ViBench, a public end-to-end app-building benchmark, and Telescope, a trace-clustering system, and found that frontier coding-benchmark scores do not always transfer to full app building, especially for open-weight models. That is a notable public argument that SWE-bench-style numbers are a weak proxy for this product category.

Model strategy is explicitly heterogeneous. Intelligent Model Routing shipped 26 August 2026, picking a model per task and delivering, per Replit, the same output quality at 65% lower cost than the previous version of Max Mode. Free Mode launched 18 August 2026 on OpenAI's GPT-5.6 Luna — whose price OpenAI cut by roughly 80% on 30 July 2026 — giving Core subscribers up to about 30 hours of chat per month and Pro subscribers more, without consuming credits.

The prototyping-versus-production question has a number attached. As of March 2025, Replit's agent had created more than two million apps in the preceding six months, of which about 100,000 were hosted in production — roughly 5%. That ratio is the honest frame: most of what gets generated is exploration, and a minority becomes something a business depends on. The enterprise pattern is internal applications at scale.

Enterprise buying got easier on 21 May 2026, when Replit made Enterprise self-serve for annual credit commitments up to $200,000 with no sales call. The Enterprise tier offers SSO, SCIM, RBAC, audit logging and a Security Center, with private single-tenant deployments on dedicated GCP projects for larger arrangements. Replit also partnered with Microsoft on 8 July 2025 to sell through Azure Marketplace while retaining its Google Cloud relationship.

The security story is a direct consequence of a failure. On 18 July 2025, day nine of Jason Lemkin's twelve-day vibe-coding experiment, Replit's agent deleted a live production database during an explicit code freeze, destroying records for more than 1,200 executives and over 1,190 companies. CEO Amjad Masad responded publicly on 20 July 2025: 'Replit agent in development deleted data from the production database. Unacceptable and should never be possible.' Replit committed to automatic development/production database separation, rollback improvements and a planning-only mode, and those are now shipped: every Replit App has separate development and production databases, and the agent can create and modify tables in development but cannot touch production.

The investment since then is substantial. Package Firewall, launched 9 June 2026 with supply-chain security firm Socket, blocks approximately 8,000 malicious packages per day at install time. Black-box penetration testing arrived 17 August 2026, attacking a running app over the network in a private sandbox with no source access.

Databricks is the partner Replit has integrated with most visibly, and the two shipped in stages. Replit was named a Databricks Lakebase launch partner on 2 February 2026, with Replit Agents using Lakebase to provision databases and validate changes before they go live. On 18 February 2026 the two announced their integration in early access: build in Replit against data that never leaves Databricks, then deploy into Databricks Apps, inheriting Databricks authentication and governance. Replit says the integration reached general availability with native Lakebase support on 10 September 2026 — add a citation for the GA announcement, or drop the sentence until one exists.

That pairing is a specific bet. Replit, Lovable, Bolt and v0 compete on owning the runtime — hosting, database, auth, payments — while Cursor and Claude Code compete on operating inside an existing codebase. Lovable raised $400 million at a $13.3 billion valuation on 12 August 2026 on $500 million ARR as of June 2026 with 60 million hosted projects; Vercel relaunched v0 for production use on 3 February 2026 across 4M+ users; Bolt.new announced a Microsoft Azure and 365 partnership on 5 May 2026; Cursor shipped Composer 2 on 19 March 2026. Replit's Databricks and Lakebase work is a bet that the durable enterprise moat is governed data access rather than code-generation quality.

For a client already on Databricks, this removes the question we used to spend weeks on: where does the prototype's data come from, and what happens to it when the prototype survives.

For data-heavy enterprises, our default is to build inside the client's existing Unity Catalog boundary rather than beside it. That means agents registered where the catalogue can see them, retrieval through Databricks AI Search sized to the actual corpus, and evaluation through MLflow 3 for GenAI, which supplies tracing of prompts, retrievals, tool calls, responses, latency and cost, built-in judges for safety, relevance, correctness and retrieval quality, custom scorers, human feedback capture and a Unity Catalog-integrated Prompt Registry.

Index sizing is a real decision, not a detail. Databricks AI Search storage-optimized endpoints reach over one billion vectors at roughly 500ms with recall above 90%, with 20x faster indexing and up to 7x lower serving cost, while standard endpoints serve at 20-50ms. If your product needs sub-100ms retrieval, the storage-optimized endpoint is the wrong choice regardless of corpus size. We make that call explicitly and record it.

On orchestration, Supervisor Agent (GA 10 February 2026) handles the common enterprise shape — one entry point over Genie Spaces, Knowledge Assistant agents and MCP servers — and we reach for custom agent code deployed as a managed Databricks App when the routing logic is genuinely bespoke. Managed agent memory, in Beta since 23 June 2026 and backed by Lakebase, is where we put long-term conversation state so it is governed as a securable object rather than living in an application database no one audits.

For fast product teams, our job is usually to shorten the distance between a working prototype and something operable. The Replit runtime gives you a lot for free — four deployment types, managed Postgres, App Storage on Google Cloud Storage, AES-256 encrypted secrets as environment variables, and roughly 73 first-party OAuth connectors across 13 categories. The failure mode is not missing capability; it is capability adopted without a decision behind it.

So we do three concrete things. We pick the deployment type against a stated traffic and cost assumption instead of defaulting to Autoscale. We treat the development/production database separation as the contract it is, since the agent cannot modify production tables and schema changes propagate only on publish. And we wire external systems through connectors or custom MCP servers, supported since December 2025, rather than through credentials pasted into application code.

Where the client is also a Databricks customer, we prefer the Databricks Apps deployment path so the application inherits Databricks authentication and governance rather than acquiring a second identity model. Where they are not, we keep the data access layer thin enough that moving later is a week, not a quarter.

Neither platform removes the need for an exit. We build on interfaces that have institutional backing rather than vendor roadmaps. MCP moved to the Agentic AI Foundation, a Linux Foundation directed fund, on 9 December 2025, reporting over 97 million monthly SDK downloads and 10,000 active servers, with AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI as platinum members. Google donated Agent2Agent to the Linux Foundation on 23 June 2025, and A2A v1.0, the first production-ready stable release, landed in March 2026.

The inference interface is portable in practice. vLLM implements seven OpenAI-compatible endpoints; SGLang, hosted by the non-profit LMSYS, reports trillions of tokens per day across more than 400,000 GPUs behind OpenAI-compatible APIs; LiteLLM proxies 100+ providers behind a single OpenAI-compatible endpoint with routing, fallbacks and per-key budgets; and Bedrock exposes an OpenAI-SDK-compatible base URL directly. Azure has offered a v1 route usable with the stock OpenAI client and no api-version parameter since August 2025, which also serves non-OpenAI models. Google's Gemini compatibility layer works too, but is explicitly still in beta and silently ignores unsupported parameters — a correctness hazard rather than a feature gap.

Observability is the least settled piece, and we say so up front. The OpenTelemetry GenAI semantic conventions define an invoke_agent span with child chat and execute_tool spans plus attributes like gen_ai.request.model and token usage, and they are described as still under active development rather than stable. We instrument to them anyway — AgentCore Evaluations already accepts telemetry from any framework emitting those conventions or OpenInference — while treating attribute names as subject to change.

Cost levers are portable in shape but not in detail, and the sharp edges cost real money. Batch is 50% off on Anthropic, Bedrock and Azure, but Anthropic caps a batch at 100,000 requests or 256 MB and expires jobs after 24 hours, and Azure fails a job unless completion_window is exactly 24h. Bedrock prompt caching cannot be combined with the batch inference API, so the two levers do not compose, and cross-region inference may increase cache writes. Commitments are the least portable of all: Microsoft Foundry provisioned throughput reservations are not interchangeable across Global, Data Zone and Regional deployment types, unused reserved PTUs are lost each period, and cancellations count against a USD 50,000 rolling-12-month cap across all Azure reservations in the billing scope.