Agent Architecture: Multi-agent: Detection Agent (analyses transaction graph patterns, velocity, geo anomalies) → Enrichment Agent (fetches merchant info, past chargebacks, device fingerprint) → Decision Agent (approve/flag/block with confidence score) → Explanation Agent (writes human-readable reason for compliance). Uses pgvector for historical pattern similarity, Redis for real-time velocity limits.

Agent Architecture: Agent pipeline: Data Fetch Agent (pulls 90 days of transactions via Plaid MCP) → Categorisation Agent (LLM labels and normalises each transaction) → Insight Agent (compares to benchmarks, identifies anomalies, flags subscriptions) → Report Agent (personalised weekly email). Episodic memory stores past months for trend analysis.

Agent Architecture: Supervisor agent routes: Filing Agent (fetches + parses SEC EDGAR 10-K/10-Q) → Earnings Agent (transcribes and summarises earnings calls) → News Agent (aggregates and sentiment-scores last 30 days) → Analyst Agent (synthesises into structured investment brief with bull/bear cases). Output: Markdown + PDF brief in under 5 minutes.

Agent Architecture: Agent workflow with HITL: Document Agent (OCR + extract income, assets, debts from uploaded docs) → Verification Agent (cross-references with credit bureaus, public records) → Risk Agent (runs DCR, LTV, stress tests) → Draft Agent (writes credit memo in standard format) → Human Underwriter reviews draft (interrupt node). The human approves, modifies, or rejects. Full audit trail via PostgresSaver.

Agent Architecture: Read-only analysis agent: Portfolio Agent (fetches current holdings) → Drift Agent (calculates deviation from target) → Tax Agent (computes tax impact of rebalancing options) → Recommendation Agent (proposes optimal rebalance minimising tax). Always read-only and recommends only — actual trades require explicit user action in the app. No write access to brokerage ever.

Agent Architecture: Pipeline agent: OCR Agent (extracts data from receipts/invoices) → Categorisation Agent (classifies expense type, maps to chart of accounts) → Policy Agent (checks against company expense policy — is this merchant category allowed? Is amount within limit?) → Anomaly Agent (flags outliers vs employee's historical patterns) → Approval routing. Integrates with Netsuite/SAP via MCP.

Agent Architecture: Scheduled agent (runs every 15 min): Rate Fetch Agent (pulls live FX rates from multiple providers) → Historical Agent (retrieves 90-day rate history from TimescaleDB) → Analysis Agent (calculates percentile rank of current rate — "This EUR/USD rate is in the 87th percentile of the last 90 days") → Notification Agent (sends Slack/email when rate crosses user-defined threshold). Stores all rates in TimescaleDB for trend analysis.

Agent Architecture: Document processing pipeline: Extraction Agent (OCR + structure extraction from Aadhaar, PAN, passport) → Validation Agent (checks document authenticity signals, expiry, consistency between documents) → Cross-reference Agent (matches name/DOB across all submitted documents) → Risk Score Agent (produces AML risk score based on PEP lists, adverse media) → Human Review queue for edge cases. Integrates with DigiLocker API for verification.

Agent Architecture: Multi-step agent: Case Intake Agent (parses dispute claim, categorises reason code) → Evidence Agent (fetches transaction data, merchant response, shipping proof, customer communication history) → Analysis Agent (compares evidence against Visa/MC dispute reason code rules) → Draft Decision Agent (writes preliminary ruling with evidence summary) → Compliance Agent (checks ruling against regulation E, zero-liability policy). Human arbiter reviews final ruling.

Agent Architecture: Batch processing agent: For each company in portfolio → Scraper Agent (fetches sustainability reports, CSR pages, regulatory filings) → Scorer Agent (extracts ESG data points, scores against MSCI/Sustainalytics methodology) → Portfolio Agent (aggregates company scores into portfolio-level ESG metrics) → Report Agent (generates quarterly ESG report with SFDR disclosures). Stores all data in Neo4j for supply chain ESG relationships.

Agent Architecture: Continuous agent: Competitor Agent (scrapes competitor prices for top 10K SKUs every 2 hours) → Demand Agent (queries sales velocity from ClickHouse, builds demand elasticity estimates) → Inventory Agent (checks warehouse levels, days-of-supply, expiry dates) → Pricing Agent (recommends price adjustments maximising revenue × margin) → Approval Agent (flags >10% changes for human review, auto-applies small adjustments). A/B testing framework built in.

Agent Architecture: Conversational agent with persistent memory: Style Profiler (learns preferences from conversation + past orders) → Trend Agent (fetches current fashion trends from runway, social media) → Inventory Search Agent (finds items matching style profile via multimodal RAG — text + image embeddings) → Size Predictor Agent (uses past returns and brand sizing data to recommend size) → Outfit Composer (builds complete look combinations). Memory persists style profile across sessions in Mem0.

Agent Architecture: Exception handling agent: Detection Agent (monitors shipment status stream, flags exceptions) → Investigation Agent (fetches route history, driver notes, customs status, weather data for affected region) → Resolution Agent (determines best resolution: redeliver, reroute, refund, replacement) → Communication Agent (writes personalised customer update email/SMS with resolution timeline) → Escalation Agent (routes complex cases to human agents with full context pre-filled). Target metric: reduce manual handle time while preserving escalation quality.

Agent Architecture: Continuous agent: Aggregation Agent (pulls reviews from Google, Trustpilot, Amazon, App Store via APIs) → Sentiment Agent (classifies sentiment, extracts specific issues: delivery, quality, support) → Trend Agent (identifies emerging issues — "packaging complaints up 40% this week") → Response Agent (drafts personalised responses for negative reviews matching brand voice) → Alert Agent (flags urgent issues: safety complaints, viral negative posts) → Product Insight Agent (weekly digest to product team).

Agent Architecture: Daily batch agent: Demand Forecast Agent (queries ClickHouse for 52-week sales history, applies seasonal decomposition) → External Signal Agent (checks local events calendar, weather forecasts, holidays for each store region) → Stock Level Agent (fetches current inventory from WMS via MCP) → Replenishment Agent (calculates optimal order quantities per SKU per store) → PO Agent (generates purchase orders for supplier approval). Human buyer reviews high-impact changes; low-risk suggestions can be auto-approved only after policy and confidence checks.

Agent Architecture: Batch pipeline agent: Image Analysis Agent (uses vision model to extract product attributes from images) → Web Research Agent (searches web for similar products, extracts specifications) → Category Agent (maps product to correct taxonomy) → Content Agent (writes SEO-optimised title, description, bullet points matching brand voice) → Compliance Agent (checks for required safety disclosures for category). 50 products/minute throughput.

Agent Architecture: Daily optimisation agent: Performance Agent (fetches campaign metrics from Google Ads / Meta MCP: impressions, CTR, ROAS, CPA by campaign/ad group) → Competitor Agent (analyses auction insights, competitive positioning) → Budget Agent (identifies overspending low-ROAS campaigns and underfunded high-ROAS ones) → Copy Agent (generates new ad copy variants for underperforming ads) → Recommendation Agent (produces action plan with projected impact). Human media buyer approves changes before implementation.

Agent Architecture: Real-time agent triggered per out-of-stock event: Product Understanding Agent (reads original item attributes: brand, size, organic status, allergens, dietary flags) → Inventory Search Agent (finds all in-stock alternatives in same category within 10% price range) → Preference Agent (retrieves customer past orders and substitution history from Mem0) → Ranking Agent (scores alternatives on attribute similarity × customer preference match) → Notification Agent (sends customer top 3 options for approval within 30 seconds).

Agent Architecture: Semantic matchmaking agent: Requirements Parser Agent (extracts structured requirements from buyer RFQ document) → Supplier Search Agent (vector similarity search over supplier capability profiles in Qdrant) → Verification Agent (checks certifications, past orders, quality ratings in supplier database) → Ranking Agent (scores suppliers on capability fit × capacity × past performance × price range) → Outreach Agent (drafts personalised RFQ emails to top 5 suppliers). Uses pgvector for semantic capability matching.

Agent Architecture: End-to-end returns agent: Eligibility Agent (checks return window, purchase verification, item category policy) → Condition Assessment Agent (uses submitted photos + vision model to assess item condition) → Decision Agent (determines refund amount based on condition, policy rules) → Inventory Agent (routes item: back to stock, return to vendor, liquidation, dispose) → Communication Agent (generates customer notification with resolution and timeline) → Finance Agent (triggers refund in payment system with appropriate amount).

Agent Architecture: LangGraph workflow triggered by GitHub webhook: Code Analysis Agent (reads diff, identifies changed logic, flags complexity) → Security Agent (runs Semgrep/Bandit patterns, checks for common vulnerabilities) → Test Coverage Agent (identifies new code paths without test coverage) → Style Agent (checks against team conventions, linting rules) → Summary Agent (writes concise review with critical issues highlighted) → Human Approval interrupt → Post Agent (posts review to GitHub). PostgresSaver for full audit.

Agent Architecture: Incident response agent triggered on alert: Context Agent (fetches alert details, affected service, recent deploys from Datadog/Sentry MCP) → Log Analysis Agent (searches Elasticsearch for error patterns in the 30 minutes before alert) → Dependency Agent (checks health of upstream/downstream services) → Runbook Agent (semantic search over past incident runbooks to find similar incidents and their resolutions) → Diagnosis Agent (synthesises all context into: likely cause, affected customers, recommended actions, escalation path) → Notification Agent (posts to incident Slack channel with full briefing). Target metric: reduce mean time to triage and improve first-response context.

Agent Architecture: Pre-planning agent: Velocity Agent (fetches last 6 sprints from Jira MCP, calculates team velocity per engineer) → Backlog Agent (reads all candidate tickets, extracts effort estimates, dependencies) → Capacity Agent (reads upcoming PTO, holidays, on-call rotations) → Dependency Graph Agent (builds dependency graph in Neo4j, identifies critical path) → Recommendation Agent (proposes sprint backlog that maximises value delivery within capacity constraints) → Output: a proposed sprint plan document for team review and modification.

Agent Architecture: Continuous documentation agent: Change Detection Agent (monitors GitHub for PRs merged to main, identifies significant changes) → Impact Agent (determines which documentation pages are likely affected by code changes) → Draft Agent (reads changed code + current docs, writes updated documentation sections) → Review Agent (flags updates for owner approval) → Search Agent (semantic Q&A over all docs via RAG) → Freshness Agent (weekly scan flagging docs not updated in 90 days with major code changes since). All docs indexed in Qdrant for semantic search.

Agent Architecture: Continuous monitoring agent: Baseline Agent (maintains rolling 2-week performance baselines in TimescaleDB for all service endpoints) → Anomaly Agent (statistical anomaly detection — flags metrics deviating >2σ from baseline) → Correlation Agent (correlates regressions with recent deploys, config changes, traffic increases) → Root Cause Agent (digs into Datadog traces — which specific function/query degraded?) → Attribution Agent (links regression to the specific commit + PR that introduced it) → Alert Agent (notifies PR author and team lead with full context).

Agent Architecture: Triage and remediation agent: Scanner Agent (runs Snyk/npm audit, collects all vulnerability alerts) → Severity Agent (fetches CVSS scores, exploitability data, whether exploit code is public) → Exposure Agent (analyses which vulnerable packages are actually used in production code paths via static analysis) → Priority Agent (ranks by: severity × exploitability × production exposure) → Remediation Agent (reads package changelog, generates the exact upgrade diff needed, checks for breaking changes) → PR Agent (opens GitHub PR with upgrade + test results). Top 20 vulns remediated as PRs, rest as prioritised backlog.

Agent Architecture: Weekly cost analysis agent: Inventory Agent (fetches all AWS/GCP resources via cloud APIs) → Utilisation Agent (queries CloudWatch/Stackdriver for CPU, memory, network utilisation over 30 days) → Waste Detection Agent (identifies: idle instances <5% CPU, over-provisioned DBs, unused reserved instances, orphaned snapshots) → Savings Agent (calculates monthly savings from each action) → Recommendation Report (ranked by savings) → Terraform diff generation for approved changes.

Agent Architecture: IaC review agent triggered on PR: Parser Agent (reads Terraform plan output, identifies all resource changes) → Security Agent (checks IAM policies for over-privilege, security groups for open ports, encryption settings) → Cost Agent (estimates monthly cost delta from resource changes) → Drift Agent (compares intended state vs actual cloud state using Terraform state) → Blast Radius Agent (identifies what would be affected if this change fails) → Review Report (structured findings with severity, remediation suggestions). Integrates with OPA for policy checks.

Agent Architecture: Continuous monitoring agent: Collector Agent (fetches CI build results from GitHub Actions API, stores test pass/fail per run in TimescaleDB) → Flakiness Agent (identifies tests with 5-30% failure rate across 30+ runs — true flakiness vs consistent failure) → Pattern Agent (correlates flakiness with: time of day, parallel job count, specific changed files, infrastructure events) → Root Cause Agent (reads the failing test code + error messages, hypothesises cause) → Fix Agent (proposes specific code fix) → PR Agent (opens PR with fix + explanation).

Agent Architecture: NL-to-query agent: Intent Agent (understands the observability question: "what services have the highest error rate today?") → Schema Agent (fetches available metrics, log indices, dashboard list from Grafana API) → Query Generator Agent (translates intent to appropriate query: PromQL for metrics, DSL for logs, SQL for traces) → Execution Agent (runs query against Prometheus/Elastic with rate limit guard) → Visualisation Agent (generates chart config or table from results) → Explanation Agent (explains what the data shows in plain English).

Agent Architecture: Pre-visit summary agent: Records Agent (fetches patient notes, labs, medications, diagnoses from Epic FHIR API — read-only) → Chronological Agent (orders events in medical timeline) → Summary Agent (synthesises into structured clinical summary: active conditions, current medications, recent labs, allergies) → Alert Agent (flags critical lab values, interactions) → Formatting Agent. PHI stays within HIPAA-compliant environment.

Agent Architecture: Real-time interaction checking agent: Medication Parser Agent (extracts structured medication list with dosages from prescription image or text) → Interaction Database Agent (queries DrugBank, RxNorm, FDA APIs for known interactions between all medication pairs) → Severity Agent (classifies each interaction: contraindicated / major / moderate / minor) → Clinical Context Agent (assesses interaction risk in context of patient's conditions) → Alert Agent (presents ranked interactions with clinical significance and management recommendations to pharmacist).

Agent Architecture: Literature monitoring agent: Indexing Agent (weekly crawl of PubMed, bioRxiv, medRxiv — fetches new papers matching researcher's interest profile) → Embedding Agent (embeds abstracts + full text using BioBERT, stores in Qdrant) → Relevance Agent (semantic search over new papers against researcher's query) → Synthesis Agent (reads top 10 relevant papers, synthesises key findings) → Citation Agent (formats references). Personalised weekly digest.

Agent Architecture: Personalised coaching agent with long-term memory: Data Collector Agent (fetches steps, sleep, HRV, workout data from Health APIs) → Trend Agent (analyses 90-day trends, identifies correlations: sleep quality vs workout intensity, stress indicators) → Benchmark Agent (compares to peer cohort and personal historical bests) → Insight Agent (identifies the 2-3 most actionable insights this week) → Coaching Agent (writes personalised recommendations matching user's goals and constraints) → Memory (stores health milestones, past recommendations in Mem0).

Agent Architecture: Patient-trial matching agent: Profile Agent (structured extraction from patient record: diagnosis, biomarkers, prior treatments, lab values) → Trial Discovery Agent (searches ClinicalTrials.gov API for trials matching cancer type, biomarker, geography) → Eligibility Agent (for each trial, checks all inclusion/exclusion criteria against patient profile) → Ranking Agent (scores trials by: scientific match + patient convenience + trial phase) → Report Agent (generates patient-friendly trial summary + physician-facing detailed eligibility analysis). De-identified profiles only.

Agent Architecture: Contract analysis agent: Extraction Agent (reads contract, extracts all key clauses: payment terms, termination rights, liability, IP ownership) → Benchmark Agent (compares each clause to company's standard positions and market norms) → Risk Agent (flags non-standard terms, assigns risk level: high/medium/low, explains the business implication in plain English) → Redline Agent (proposes specific redline language for high-risk clauses) → Summary Report (executive summary + clause-by-clause risk table).

Agent Architecture: Sales intelligence agent: Trigger Agent (monitors LinkedIn, news for prospect company events: new CTO, Series B raised) → Research Agent (builds prospect profile: tech stack, headcount, competitors) → Pain Point Agent (reads prospect's job postings, engineering blog — infers problems) → Personalisation Agent (matches pain points to product capabilities) → Outreach Agent (writes 3-variant email) → CRM Agent (updates HubSpot + queues email for approval).

Agent Architecture: Personalised learning agent: Assessment Agent (identifies current skill level via quiz + past course completions) → Gap Agent (compares current skills to target role competencies) → Curriculum Agent (searches course library semantically — "Python for data engineers" — matches gaps) → Path Builder Agent (sequences courses optimally) → Progress Agent (weekly check-in, adjusts path) → Manager Report Agent. Memory tracks learning journeys.

Agent Architecture: L1 resolution agent: Intent Agent (classifies ticket type from user description + attachment) → Knowledge Agent (searches IT runbook database via RAG for resolution procedure) → Eligibility Agent (checks if user is authorised for requested access via Active Directory) → Resolution Agent (executes resolution: resets password via AD API, provisions software via JAMF) → Communication Agent → Escalation Agent (routes unresolved to L2). Resolves 65% of tickets autonomously.

Agent Architecture: Full-cycle recruiting agent: JD Analysis Agent (extracts requirements from JD) → Sourcing Agent (searches LinkedIn, GitHub for candidates matching technical criteria — active and passive) → Screen Agent (compares candidate profile against JD requirements, produces structured scorecard) → Outreach Agent (writes personalised recruiter message) → Interview Prep Agent → Feedback Agent. Decisions audited for bias.

Agent Architecture: Automated eval agent: Benchmark Agent (runs standard benchmarks: MMLU, HumanEval on new model version) → Regression Agent (compares scores to previous version, flags regressions) → Red-Team Agent (generates adversarial prompts targeting known failure modes: jailbreaks, hallucinations, biases) → Comparison Agent (A/B tests new vs old model, human-judges outputs) → Report Agent (produces capability vs safety scorecard).

Agent Architecture: Deep research agent: Query Planning Agent (decomposes research question into 8-15 specific sub-questions) → Parallel Search Agents (5 parallel search agents, each handling 2-3 sub-questions with Tavily) → Synthesis Agent (reads search results, identifies key themes, contradictions) → Fact-Check Agent (cross-validates key claims across multiple sources) → Writer Agent (produces 2,000-word structured report with inline citations) → Citation Agent (formats source URLs). Uses streaming.

Agent Architecture: Tiered moderation agent: Classifier Agent (fast first-pass classification: clear violation / borderline / clearly safe) → Context Agent (for borderline cases: fetches user history, channel context) → Nuance Agent (analyses borderline content with full context — satire, news) → Policy Agent (maps violation to specific policy section with justification) → Action Agent (determines action: remove, age-gate, add warning) → Appeal Agent. Humans review only borderline cases.

Agent Architecture: Continuous fraud monitoring agent: Telemetry Agent (ingests GPS trace, speed, acceleration data from trips in ClickHouse) → Anomaly Agent (statistical detection — trips with impossible speeds, GPS spoofing signatures) → Network Agent (graph analysis in Neo4j — identifies clusters of accounts that frequently interact) → Incentive Abuse Agent (checks bonus claim patterns) → Risk Scorer Agent → Human Review queue. Real-time trip scoring.

Agent Architecture: Listing optimisation agent: Property Analyser Agent (reads current listing, photos via vision model) → Competitive Agent (searches similar listings in same area, identifies differences) → SEO Agent (analyses search ranking factors, identifies keyword opportunities) → Copywriter Agent (rewrites title, description, highlight bullets) → Pricing Agent (analyses demand, events, competitor pricing to recommend dynamic pricing) → Photo Audit Agent.

Agent Architecture: Predictive maintenance agent: Sensor Agent (ingests vibration, temperature, pressure from IoT sensors via MQTT into InfluxDB) → Anomaly Agent (detects deviation from normal operating envelope) → Trend Agent (identifies gradual degradation patterns) → RUL Agent (Remaining Useful Life prediction — "Bearing #3 on Compressor A: estimated 14 days before failure") → Work Order Agent (creates work order in SAP with parts) → Outcome Agent.

Agent Architecture: Real-time news agent: Monitor Agent (subscribes to SEC filings, earnings wires — triggers on new documents) → Parser Agent (extracts structured data: EPS, revenue, quotes) → Context Agent (fetches historical context: same period last year, consensus, stock reaction) → Writer Agent (generates news article in AP/Reuters style) → Fact-Check Agent (verifies numbers against source) → Editor Agent (checks style, legal concerns).

Agent Architecture: Personalised tutor agent: Assessment Agent (analyses exercise performance to identify specific weaknesses) → Curriculum Agent (dynamically adjusts lesson sequence) → Content Generator Agent (creates novel practice sentences using learner's interests and vocabulary) → Conversation Partner Agent (conducts spoken dialogue practice, corrects errors) → Progress Agent (tracks retention via spaced repetition).

Agent Architecture: End-to-end RFQ processing agent: Document Agent (OCR + extract specs from drawings, BOMs) → Specification Agent (structures requirements: material, tolerances, quantity) → Capability Check Agent (verifies factory can meet specs) → Material Agent (fetches prices, calculates cost) → Process Agent (estimates machining hours based on geometry) → Pricing Agent (applies overhead, margin) → Quote Agent (generates quote). Human reviews before sending.

Agent Architecture: Threat intelligence agent: Ingestion Agent (pulls IOCs from VirusTotal, MISP, Shodan — IPs, hashes) → Correlation Agent (uses Neo4j to map relationships: IP uses same ASN as C2, same TLS cert) → ATT&CK Agent (maps observed TTPs to MITRE ATT&CK actor profiles) → Priority Agent (scores threats by exposure, exploit availability) → Report Agent (generates daily brief with recommended detections) → SIEM Agent (pushes IOCs to SIEM for automatic blocking).