01 · Cloud & Hosting close ✕
Cloudflare A global compute fabric where stateful agents, vector memory, model inference, browser automation, and a model-agnostic gateway live behind one binding model.
AI / agentic features Agents SDK. Durable-Object-backed agents with isolated SQLite, WebSockets, scheduling, and resumable streams. Workers AI. Inference at 300+ POPs with a unified binding to first-party + 14+ third-party model providers. AI Gateway. Caching, rate-limit, retries, model fallbacks, OTel logs, unified billing. Vectorize + AI Search. Vector DB plus a hybrid BM25 + semantic primitive in one query. Browser Run + Durable Object Facets. Browser-as-a-service with Live View and human-in-the-loop; per-tenant SQLite DBs spun up on demand. How we'd wire this into an ABOS Cloudflare is the default substrate for greenfield agent work. Workers + Durable Objects + Vectorize + Agents SDK + AI Gateway collapse what would be Lambda + DynamoDB + OpenSearch + Step Functions + a separate gateway on AWS into one platform with one binding model. AI Gateway emits OTel for the same observability spine that records token cost, p95, and safety-harness pass rate. The "owned, not rented" property at the model layer.
01 · Cloud & Hosting close ✕
AWS The default enterprise substrate when the customer already lives in AWS. Bedrock fronts the model layer. AgentCore is the production agent runtime. Guardrails plus Knowledge Bases give safety and RAG without building them.
AI / agentic features Bedrock Agents. Managed agent runtime with tool use, S3-backed Knowledge Bases, code interpreter, supervisor-pattern multi-agent collaboration. Bedrock AgentCore. MicroVM-isolated agent platform with Runtime, Memory, Identity, Gateway, Observability primitives. April 2026 added Managed Harness, AgentCore CLI, Filesystem Persistence, A2A protocol. Bedrock Guardrails. Content filters, PII redaction, Automated Reasoning checks. ApplyGuardrail works on non-Bedrock models too. Bedrock Knowledge Bases. Managed RAG with hybrid search and embedded retrieval-time guardrails. SageMaker Unified Studio. Single workspace combining Bedrock Knowledge Bases, Agents, Guardrails, Flows, plus Nova Forge SDK. How we'd wire this into an ABOS Picked when data gravity, compliance posture (FedRAMP, HIPAA BAA, regulated PE portfolios), or existing AWS spend make leaving cost-prohibitive. AgentCore Runtime hosts the workflow agents (quote-approval/auto-route, case/triage-summarize); Knowledge Bases back the corpus; Guardrails wire into the deterministic pass-rate metric on the observability spine. Heavier per workflow than Cloudflare; right answer when the customer's CFO needs to see "AWS" on the architecture diagram.
01 · Cloud & Hosting close ✕
Microsoft Azure The substrate when the customer's center of gravity is Microsoft 365, Dynamics, or Fabric. Foundry Agent Service is the agent runtime. Copilot Studio is the business-user composition layer on top.
AI / agentic features Foundry Agent Service. GA March 2026; built on the OpenAI Responses API, wire-compatible with OpenAI Agents, supports DeepSeek, xAI, Meta, plus LangChain/LangGraph; private networking, Entra RBAC, full tracing. Connected Agents. Agents call other agents as tools for multi-agent orchestration. · preview Azure OpenAI Service. First-party hosting of OpenAI models with enterprise data-residency and Entra-integrated access. Foundry tracing + evaluations. GA tracing UX with sort/filter and eval-results-to-trace linking. Copilot Studio. Low-code agent composition that publishes into Microsoft 365 surfaces (Teams, Outlook). How we'd wire this into an ABOS The substrate when ABOS workflows need to land inside Microsoft surfaces (Teams-resident triage agent, Outlook-driven approval, Dynamics-record automation) or when private-networking and Entra RBAC are non-negotiable. Connected Agents gives the supervisor pattern; tracing-to-eval linking maps directly onto Elevate's deterministic safety-harness pass-rate metric.
01 · Cloud & Hosting close ✕
Google Cloud Google's consolidated agent platform (Vertex AI Agent Builder plus Agentspace, rebranded and unified at Cloud Next 2026). Strongest when BigQuery is the data layer and the corpus is Workspace-native.
AI / agentic features Gemini Enterprise Agent Platform. Unified successor to Vertex AI Agent Builder + Agentspace, announced at Google Cloud Next 2026. Agent Studio + Agent Development Kit. No-code design + code-first SDK for agent composition. Agent Garden. Library of prebuilt agents and templates. Agent Runtime. Managed runtime with sub-second cold starts and long-running agent support. Sessions + Memory Bank + A2A protocol. Stateful sessions ($0.25 per 1,000 events), shared memory bank, open agent-to-agent interop standard. How we'd wire this into an ABOS Picked when the customer's analytical center is BigQuery, when Workspace is the primary corpus surface, or when Gemini's long-context window genuinely outperforms alternatives on the workload. ADK + Agent Runtime cover the same ground as AgentCore but with tighter Google-data integration.
02 · Data & Warehousing close ✕
Snowflake Governed warehouse with first-party LLM inference, hybrid retrieval, and declarative agents that run inside the customer's Snowflake account.
AI / agentic features Cortex AI Functions. Call LLMs (`AI_COMPLETE`, `AI_CLASSIFY`, `AI_FILTER`, `AI_AGG`) directly from SQL. Cortex Search. Managed hybrid (semantic + keyword) retrieval over docs and tables. Cortex Analyst + Semantic Views. Text-to-SQL grounded in YAML/Semantic-View definitions; Semantic View Autopilot GA Feb 3, 2026. Cortex Agents. Declarative agents orchestrating Cortex Analyst + Cortex Search, with native MCP support and Brave web-search tool calls. Document AI + Snowpark Container Services. Extract structured fields from PDFs/images; run custom containers (incl. inference) inside Snowflake's governance perimeter. How we'd wire this into an ABOS Strongest fit when the customer already lives in Snowflake. Cortex Agents map cleanly onto quote-approval/auto-route and case/triage-summarize patterns: Semantic Views encode the business model, Cortex Search handles the corpus, Cortex Agents stitch the two and call MCP tools. All embeddings, inference, agent orchestration run against governed tables without exfiltration. Watch out: per-query Cortex costs can spike; the observability spine should track Cortex token + warehouse-credit cost per workflow execution.
02 · Data & Warehousing close ✕
Databricks Lakehouse with end-to-end agent tooling (Mosaic AI), governed by Unity Catalog and queryable in natural language via Genie.
AI / agentic features Mosaic AI Vector Search. Managed vector index over Delta tables with built-in retrieval-quality eval. Agent Bricks. Describe-the-task agent builder with auto-generated evals + cost/quality tuning. MCP-native. AI/BI Genie + Genie Code. Text-to-SQL/code grounded in Unity Catalog Business Semantics with ABAC, row filters, column masks. ai_parse_document + ai_prep_search SQL functions. Parse PDFs/Office docs, chunk, and write to a vector index entirely in SQL. Mosaic AI Model Serving + AI Gateway. Unified endpoint for foundation + custom models with usage logging and policy controls. How we'd wire this into an ABOS Right pick when the client's analytics center of gravity is already a lakehouse, or when the workflow needs heavy custom-model fine-tuning (Mosaic AI's strength). Unity Catalog / Genie / Agent Bricks map onto Elevate's "deterministic safety harness" framing. Metric Views are the contract the agent must honor, Genie is the dashboard surface, Agent Bricks is where opportunity/intent-classify lives. Like Snowflake, inference runs inside the customer's workspace.
02 · Data & Warehousing close ✕
Postgres + pgvector The pragmatic, OLTP-shaped data tier. Operational tables and vectors in one Postgres, no warehouse bill.
AI / agentic features pgvector 0.8.x. HNSW + IVFFlat indexes, iterative scans for filtered queries, parallel HNSW build, halfvec quantization. Hosted variants. Supabase, Neon, AWS RDS, Google Cloud SQL, Azure Database for PostgreSQL all ship pgvector as a first-class feature. BYO inference. Postgres queries an external model endpoint (OpenAI/Anthropic/Bedrock/Vertex) via app code or `http`/`plv8` extensions. Operational fit. Same DB holds agent transactional state and corpus, simplifying the "owned" surface. How we'd wire this into an ABOS Right when the client is mid-market with no existing warehouse, when Snowflake/Databricks spend doesn't pencil, or when the ABOS is fundamentally an application (not an analytics surface). The Salesforce-shaped record model fits Postgres natively. Honesty point: data leaves the perimeter for inference. Most mid-market workloads accept that; regulated workloads use Snowflake/Databricks instead.
02 · Data & Warehousing close ✕
Google BigQuery Serverless warehouse with native Gemini inference, embeddings, and vector search exposed as SQL functions and data-agent surfaces.
AI / agentic features AI.GENERATE / AI.GENERATE_TABLE. Multimodal text/image/audio/video/PDF generation as table-valued SQL functions. AI.EMBED. Single embedding from mixed-modality input via the latest Gemini embedding model. · preview VECTOR_SEARCH. Native vector index with single-query optimization. · preview Data Agents + Conversational Analytics. Data agents grounded in BigQuery tables/views/UDFs; integrated with Gemini Enterprise Agent Platform; MCP-enabled. How we'd wire this into an ABOS Natural fit if the client is GCP / Workspace-native. The AI.* functions keep inference inside the Google trust boundary; Conversational Analytics gives a ready-made Genie/Cortex-Analyst-equivalent surface. Slightly more wiring than Snowflake/Databricks because the agent platform is a separate product (Gemini Enterprise) rather than living inside the warehouse.
03 · AI Models close ✕
Anthropic Best-in-class tool use, native MCP, and the only major vendor offering a hosted long-horizon agent runtime with first-class checkpointing.
AI / agentic features Claude Opus, Sonnet, Haiku. Frontier family with full extended-thinking support and long-context Sonnet variants for whole-codebase / whole-corpus reasoning. Tool use + first-party tools. Parallel tool calls; built-in `web_search`, `web_fetch`, `code_execution`, `computer_use`. Model Context Protocol (MCP). Anthropic-originated open standard for agent-to-tool wiring; broadly adopted across vendors. Prompt caching. 5-min cache reads at 0.1× input price; 1-hour cache writes at 2×; workspace-level isolation. Claude Managed Agents. Hosted harness with sandboxed code exec, checkpointing, scoped credentials, end-to-end tracing. · preview How we'd wire this into an ABOS The orchestration brain across reference workflows: quote-approval/auto-route, case/triage-summarize, opportunity/intent-classify. MCP gives the ABOS a vendor-neutral wire format between agent and Salesforce/Snowflake/email tools. Exactly the "owned, not rented" property. Managed Agents' checkpointing and tracing emit per-step spans and token-cost signals that the observability spine depends on.
03 · AI Models close ✕
OpenAI The breadth platform. Widest tool ecosystem, only serious option for production voice agents, and the most mature visual agent-builder for non-engineering buyers.
AI / agentic features GPT family + gpt-realtime. Frontier reasoning + computer-use models, plus the only model production-grade for sub-second voice. Responses API. Agentic loop primitive with built-in tools, remote MCP servers, custom functions in a single request; 1M context with native compaction. AgentKit. Visual canvas + Connector Registry + ChatKit for embedded agent UIs. · beta Agents SDK. Open-source provider-agnostic orchestration with handoffs, guardrails, tracing. Prompt caching + Realtime API. 24-hour extended retention; async function calling so long tool calls don't break voice sessions. How we'd wire this into an ABOS Two specific slots: (1) Voice-first workflows (case/triage, inbound qualification, field-tech support) where gpt-realtime is the only model currently production-grade for sub-second voice. (2) AgentKit's visual canvas + Connector Registry lets us hand part of the workflow surface to a client's ops team for safe edits without touching agent code, which directly serves "owned, not rented."
03 · AI Models close ✕
Google The long-context multimodal specialist. Strongest grounding tools and the most disciplined "thought signature" reasoning trace, with a smaller agent-tooling ecosystem than the top two.
AI / agentic features Gemini Pro + Flash. `thinking_level` parameter (low/medium/high) for explicit cost-vs-reasoning trade-off. Thought Signatures. Encrypted reasoning state passed back into the next turn; required on the latest Gemini for multi-turn function calling reliability. Tool combos + context circulation. Built-in `google_search`, `google_maps` composable with custom functions; tool outputs preserved across turns. Interactions API. Newer alternative to `generateContent` for state, tool orchestration, long-running tasks. · beta Context caching + Deep Research agent. 75% discount on cached reads; collaborative planning, MCP server integration, file search. How we'd wire this into an ABOS Slots where the workflow needs heavy multimodal input (PDFs, scanned docs, images of forms) or live-grounded answers (opportunity/intent-classify where intent depends on current company news). Thought Signatures give the deterministic safety harness a more verifiable reasoning trace than free-form thinking.
03 · AI Models close ✕
Local models Open-weight model families that run inside your VPC, on bare metal, or on a self-hosted GPU pool. We pick them when data residency, regulatory perimeter, or per-token economics make hosted APIs the wrong answer.
Families we deploy from Meta
Llama
MoE and dense variants with permissive license and the broadest open-weight ecosystem; the safe default when no other constraint pulls harder.
DeepSeek
DeepSeek
MoE, MIT-licensed, strong reasoning. Cost-efficient inference; first-tier choice when reasoning matters and budget is tight.
Moonshot
Kimi
Long-context, agentic-tuned, open weights. Strong tool-use traces and tight instruction-following.
Mistral
Mistral
EU-residency default. Lighter footprint; right answer for European clients with GDPR-strict data postures.
Alibaba
Qwen
Multilingual, strong tool calling, broad size range. Solid for non-English workflows and APAC clients.
Google
Gemma
Small-to-mid open models. Pairs well with the rest of a Gemini-led stack as a cheap classifier tier.
How we'd wire this into an ABOS The per-step optimization lever. High-volume classifier nodes (case/intent-tag, quote/eligibility-check) where a frontier model is overkill, and the regulated workflows whose data cannot leave a customer VPC. Treated as a fungible tier behind the agent runtime: same prompt contracts, same eval datasets, swappable per-workflow.
04 · Business Platforms close ✕
Salesforce The dominant CRM system-of-record for B2B revenue, service, and field operations. Now the most mature agentic platform on the market after the 2025 rebrand of the "Cloud" suite into Agentforce 360.
AI / agentic features Agentforce. Autonomous agents grounded in CRM/Data Cloud taking multi-step actions on Opportunity, Case, Quote, Order. Command Center, MCP support, Agent Script (deterministic + LLM hybrid). 100+ pre-built industry actions. Atlas Reasoning Engine. Multi-step planner with RAG over Data Cloud, tool-loop, real-time response streaming. Data Cloud zero-copy federation. Query Snowflake / Databricks / BigQuery in place; consent and governance enforced at execution. Einstein Trust Layer. PII masking, zero-data-retention LLM contracts, full audit log to Data Cloud's Audit DMO. FedRAMP High. Prompt Builder · Model Builder. Versioned prompts callable from Apex/Flow; bring-your-own-model. How we'd wire this into an ABOS Natural action layer for every reference workflow. quote-approval/auto-route, opportunity/intent-classify, case/triage-summarize, order/fulfillment-route are all Salesforce-shaped object names. The agent reasons in Atlas, calls a Topic/Action via Agentforce, and the write persists on the record. Trust Layer audit data flows into the observability spine. Every prompt and outcome queryable alongside p95 and execution count.
04 · Business Platforms close ✕
ServiceNow The system-of-record for cross-functional workflow (ITSM, CSM, HRSD, SecOps), and the platform whose 2026 Zurich release positions it as the agentic collaboration backbone for the enterprise.
AI / agentic features AI Agent Studio. Natural-language authoring of custom agents; pre-built agents for ITSM, CSM, HRSD, SecOps. AI Agent Orchestrator. Coordinates multi-agent teams with handoff between specialist agents. Workflow Studio + Agentic Playbooks. Pre-built playbooks where agents handle entire cross-functional jobs end-to-end (Zurich release). AI Control Tower. Central governance hub: AI inventory, lineage, model-behavior monitoring, lifecycle orchestration. Supports governance for non-ServiceNow agents. Now Assist Guardian. Real-time guardrails for prompt injection, harmful output, sensitive-data exposure. How we'd wire this into an ABOS Where case/triage-summarize lives natively. Agent reads the Case, calls a Knowledge Graph lookup, drafts a response, writes back to the Case record through Workflow Studio actions. AI Control Tower is unusually compelling for ABOS observability. It can govern agents that don't live on ServiceNow, fitting the "owned, not rented" framing.
04 · Business Platforms close ✕
HubSpot The dominant CRM/marketing/service platform for the lower mid-market, and the most aggressive vendor on outcome-based agent pricing in 2026.
AI / agentic features Breeze Prospecting Agent. Full prospecting lifecycle: detects buying signals, builds contact lists, drafts personalized outreach, books meetings. Breeze Customer Agent. Autonomous resolution of customer conversations across chat/email/help center. Breeze Content / Social Agent. Content drafting and channel publication. Breeze Copilot + Intelligence. Role-aware in-context assistance; third-party data enrichment with standard firmographic enrichment now free. Outcome-based pricing. Customer Agent at $0.50 per resolved conversation; Prospecting Agent at $1 per lead recommended (April 14, 2026). How we'd wire this into an ABOS Hosts opportunity/intent-classify (Prospecting Agent signals → Deal records) and a customer-service equivalent of case/triage-summarize (Customer Agent → Ticket records). Less of a fit for quote-approval and order/fulfillment-route since HubSpot is weaker on quoting/order workflows than Salesforce. Outcome-based pricing translates cleanly into ABOS unit economics.
04 · Business Platforms close ✕
NetSuite Oracle's mid-market ERP system-of-record for finance, procurement, and inventory. The place ABOS-driven financial workflows have to land if the client runs NetSuite.
AI / agentic features AI Connector Service. Bring-your-own-model agent framework supporting OpenAI, Amazon Bedrock, Google Vertex AI, Microsoft Foundry (SuiteConnect 2026 R1). Custom Tool Script Type. Exposes SuiteScript logic as callable tools so agents can post journal entries, update records, trigger workflows. AI Close & Cash Management agents. Autonomous close-cycle support and EPM agents (2026.1). Bill Capture. OCR-based AP automation: vendor name, invoice number, date, line items extracted to vendor-bill records. Text Enhance + Suite Assistants. Assistive AI-generated narrative content; embedded predictions for forecasting and anomaly detection. How we'd wire this into an ABOS Where finance-side ABOS workflows persist: AP automation, three-way match, close-cycle journal entries, cash application. The Custom Tool Script Type is the key 2026 unlock: before it, NetSuite agents could read but couldn't reliably write back through a governed surface. Pairs naturally with Salesforce-led front-office workflows where the order-to-cash handoff lands in NetSuite.
05 · Observability close ✕
Langfuse OSS, OTel-native LLM engineering platform. Single source of truth for traces, evals, prompts, and datasets.
AI / agentic features OTel-native ingest. OTLP/HTTP endpoint; SDK v4 is a thin shim over the official OpenTelemetry client. Evals. Datasets, LLM-as-judge, human annotation queues, custom-pipeline scoring; thresholds wire into CI for regression gating. Prompt management. Versioned prompts pulled at runtime, SDK-side cached; A/B compare across runs. Sessions + multi-turn trace grouping. Multi-step agent runs collapse to a single inspectable trajectory. Self-host or SaaS. MIT-licensed core; enterprise add-ons (SSO, RBAC) gated. How we'd wire this into an ABOS Covers three of four Instrument tiles natively: token cost, p95 latency, workflow execution count. The fourth tile, deterministic safety harness pass rate, is Langfuse's eval module reading from a versioned dataset and posting scores back to the same trace. In the stack: trace store + eval runner in one box, exporting to SigNoz over OTel for the unified system-and-LLM view.
05 · Observability close ✕
Arize OpenInference/OTel-native eval-first observability. Phoenix (OSS, MIT) for the trace + eval core, AX (commercial) for production monitoring, drift, and CI gating at scale.
AI / agentic features OpenInference + OTel auto-instrumentation. LangChain, LlamaIndex, DSPy, Vercel AI SDK, OpenAI, Bedrock, Anthropic; seven span types including dedicated AGENT and TOOL spans. Agent trajectory mapping. Auto-detects recursive loops, repeated tool failures, dead-end branches across multi-agent graphs. Dual-level evals. Scores individual reasoning steps and final goal achievement; pluggable evaluators. Prompt management module. Versioned templates in OSS Phoenix. AX (commercial). Online evals, drift monitoring, Alyx AI assistant, CI/CD experiment gates, RBAC, HIPAA, BYOC self-host. How we'd wire this into an ABOS Same four-tile coverage as Langfuse, with stronger story for multi-agent trajectory analysis (where ABOS quote-approval-style workflows fan out). AX is the upgrade path for clients who need petabyte storage, SLAs, HIPAA. In the stack: trace store + eval runner with a clean OSS-to-enterprise ramp.
05 · Observability close ✕
LangSmith LangChain's commercial-first agent observability + evals platform. Deepest agent-trajectory and CI-eval integrations of any vendor, but framework-flavored and not OTel-native.
AI / agentic features Full agent-trajectory tracing. Every step, tool call, reasoning hop captured; framework-agnostic via Python/TS/Go/Java SDKs. Evaluation suite. Datasets, LLM-as-judge, heuristic checks, pairwise comparisons, human annotation queues. CI-integrated evals. pytest, Vitest, GitHub Actions; threshold-fail on score regression. Polly + Insights Agent. LLM assistants that summarize large traces and surface failure patterns. Deployment runtime. Managed durable runtime with exactly-once execution and human-in-the-loop approvals. How we'd wire this into an ABOS Covers all four Instrument tiles internally. Friction: not OTel-native; export to SigNoz requires a bridge, weakening the "vendor-resilient" pitch. Best-fit when an ABOS client is already standardized on LangChain/LangGraph and wants the tightest agent-debug experience available.
05 · Observability close ✕
Datadog LLM Observability Enterprise APM extended with first-class LLM tracing. The right pick when the ABOS lives inside an existing Datadog footprint.
AI / agentic features Native OTel GenAI semconv ingest. Accepts OTel 1.37+ GenAI spans without the Datadog SDK or Agent. Per-span agent tracing. Input/output, latency, token usage, errors, privacy violations on every step; auto-instrumentation for OpenAI, LangChain, Bedrock, Anthropic. External evaluations API. Push eval scores from any harness (Ragas, DeepEval, custom) into the same trace view. LLM Observability Insights. Automatic anomaly detection over rolling 7-day windows for regression / drift catch. Unified APM correlation. LLM spans link to upstream HTTP / DB / queue spans; one pane across system + agent. How we'd wire this into an ABOS Supplies cost, latency, executions natively; deterministic pass rate must come from an external eval runner pushing scores via the External Evaluations API. The right call when the client already pays for Datadog and the ABOS must merge into existing on-call dashboards and SLOs. Wrong call when budget is the constraint or when the client wants a vendor-neutral, owned trace store.
Hover or click a tile to inspect Each tile opens a panel with the platform's current AI features and how we'd wire them into an ABOS.
