Salesforce engineers describe building AI-Assisted Test Failure Buddy, a Python and FastMCP tool exposing eight MCP tools to Claude Code and Cursor, to triage over 600 test failures caused by retiring the SOAP API login mechanism. The post explains how deterministic code handles API calls, batching, signature matching, and reassignment while AI interprets unmatched stack traces and ambiguous ownership, reducing estimated two to three days of triage to about 90 minutes.
Salesforce Engineering engineering blog
Loading
Salesforce's AI Research team introduces EGI-Bench, a unified framework that consolidates eight independently developed enterprise LLM benchmarks into six skill dimensions (task outcomes, tool actioning, planning, grounding, safety, engineering) with normalized sub-metrics and a weighted global score. The post reports findings from testing 14 LLMs, including that no single model leads across all dimensions, multi-step planning remains weak, and lower-cost models perform competitively. The...
Salesforce's Hyperforce Runtime Platform engineering organization describes how it scaled AI tool adoption across roughly 500 engineers using a federated ownership model, a shared Claude Code plugin with automatic distribution, and SDLC and DORA metrics to guide decisions. The post reports measured results, including cycle time falling from 22 to 11 days and sprint delivery rising from 66% to 90%, and explains how code review and testing were extended into the inner loop.
Salesforce engineer Fereydoon Vafaei describes how the Agentforce team automates parts of the Atlas Reasoning Engine's software lifecycle. The approach combines semantic retrieval over incident history, agent-as-a-judge evaluation with human sign-off, and a risk-based model that limits autonomy to verifiable, reversible changes.
Salesforce engineers describe how they optimized Apache Spark resource sizing across eight million daily executions using listener-based telemetry, Kafka, and an Iceberg lakehouse. They explain how to distinguish unused capacity from bottlenecks, validate changes safely, and achieve a 31% reduction in targeted compute-cost run rates without significant reliability or performance regressions.
Salesforce engineers explain how Agentforce Memory evolved from a proof of concept into a production platform capability for persistent AI-agent context. The post covers shared memory terminology, semantic and keyword retrieval trade-offs, pre-orchestration, developer ergonomics, latency, and user privacy controls.
Salesforce engineers explain how deterministic orchestration makes AI-generated prompt templates reliable. The design separates LLM interpretation from graph-controlled routing, permission-aware grounding, record identity, structured output validation, and human approval.
Salesforce describes Agent Designer, a governed multi-agent system that turns engineer requests into tested agent teams in 15–30 minutes. The post explains orchestration boundaries, topology-specific budgets, structured verification, conflict reconciliation, and bounded self-healing to prevent unreliable repairs.
Salesforce engineers explain how Cloud Atlas replaced per-instance rate limiting with coordinator-free global quota management and intelligent load shedding. The approach uses queue time, priority-aware throttling, noisy-neighbor isolation, observability, and reversible rollout practices to preserve availability and prevent cascading failures in a globally distributed identity platform.
Salesforce engineers explain how Agentforce Health Monitoring detects silent AI-agent availability failures by unifying fragmented session, tool, error, and escalation telemetry. The post details streaming-ingestion and query optimizations that reduced alert latency, plus session-level drill-downs for diagnosis and a roadmap toward automated remediation.
Salesforce explains how Data 360 compiles governed, minimal context for enterprise AI agents across structured, unstructured, federated, and streaming sources. The post details a six-stage context pipeline, memory and learning controls, distributed-system architecture, and workload measurements for retrieval, concurrency, analytical SQL, and graph queries.
Salesforce engineers explain how Data 360 data graphs unify fragmented identities, products, entitlements, and account data into trusted context for AI agents. The post covers partitioned data access, graph sizing, search, reusable delivery practices, and achieving P50 latency below 200 milliseconds.
Salesforce engineers explain how they built real-time mobile personalization across iOS, Android, React Native, and Flutter. The architecture separates app instrumentation from campaign configuration, performs identity resolution and decisioning server-side, and uses approved native components, CDN-delivered metadata, and QR-based preview testing.
Salesforce engineers explain how enterprise RAG systems lose meaning across parsing, chunking, enrichment, retrieval, and generation. They present a stage-by-stage debugging method, structure-aware parsing and chunking, metadata filtering, GraphRAG, and measured improvements from roughly 46% to above 90% accuracy.
Salesforce explains how it combines configuration metadata with runtime telemetry to assess enterprise org health at scale. The post covers harmonizing fragmented data, preserving freshness and access context, and ranking deterministic findings into context-aware next-best actions.
Salesforce explains how Agentforce separates stochastic LLM reasoning from deterministic presentation using the render: directive and response formats. The architecture supports reliable component selection across multi-action, long-running, supervised, third-party, and headless experiences, with per-action audit logs for regulated use cases.
The post explains why conventional RAG can retrieve accurate but incomplete evidence for conditional decisions, leading AI agents to wrong conclusions. It shows how GraphRAG uses multi-hop knowledge-graph traversal, validated schemas, and explicit links to structured records to uncover missing relationships and diagnose retrieval failures.
Salesforce explains how its DREAM platform uses AI-driven inference and distributed orchestration to mitigate hyperscale, application-layer DDoS attacks in seconds. The post details the team’s rapid AI-assisted rewrite and production lessons, including keeping large payloads and unbounded workflow histories out of the orchestration layer.
Salesforce engineers explain how to convert machine-learning signals into actionable recommendations by combining predictions with business rules and contextual knowledge. The post also examines MCP contracts, agent ownership models, and delivering recommendations in existing workflows without creating another dashboard.
Salesforce presents a spec-driven workflow for building trust in AI-generated code. The approach makes uncertainty explicit, separates evidence-based lookups from human judgment, grounds plans in repository evidence, maps requirements to tests, and uses independent adversarial review gates.