July 23 2025 – London. Last night (the 22nd of July), inside Microsoft’s Reactor, the LangChain London crowd re-assembled—fellow developers and business creators who treat these meet-ups as impromptu peer-review sessions, bringing code and candour in equal measure (Microsoft Reactor 2025) . Marlene Mhangami opened with the Model Context Protocol, a still-forming schema for exposing a cluster of collaborating agents as one inspectable endpoint—her demo wrapped a pre-processor, a planner, and a code-executor behind a single call (Mhangami 2025; Anthropic 2024; IBM 2025) . New to me and fascinating. Pablo Elgueta Truffello argued that converting speech to text discards prosody and bungles interruptions—the very aspects that make speech richer data—and showed how full-audio pipelines, though still brittle in tool invocation, may reach production readiness within a year (Cartesia AI 2024) . Sri Rang provided the evening’s fulcrum: a compliance-first reference architecture where policy engines, audit ledgers, and jurisdiction-aware checkpoints run through the data flow itself—exactly the scaffold our BigQuery-backed agent swarm is converging on (Patel 2025; Google Cloud 2025) . That convergence confirms that ethics and regulation can be designed through, not patched on (Scott 2025) . Finally, Nuno Campos contrasted talkative, turn-based companions with ambient agents that persist, observe event streams, and enlist their peers on demand—a framing that neatly matches our own ensemble, which boots at project start, watches schema drift, and triggers collaborators the moment new meaning surfaces (Campos 2025a; Campos 2025b) .

The chapters ahead trace how our NICER-DW swarm occupies each layer of Sri Rang’s reference architecture.

• Control plane — a BigQuery catalog of versioned agent specs, typed tool stubs, and DSPy policy manifests, so scope and data-residency are locked in before the first query runs.
  
• Execution plane — LangGraph’s state-machine compiler wires three specialists under those same policies: a Reflective SQL agent whose DSPy Self-Discover programme drives tree planning, reflection, and LangMem-backed memory persisted to bigquery; a DBT Data-Vault agent that turns those semantics into AutomateDV YAML; and an Analytics agent that surfaces the warehouse’s new knowledge via the extensible and open source Apache Superset BI tool.
  
• Observability / rollback plane — every prompt, tool call, and KPI streams into LangSmith run-trees and timestamp-clustered BigQuery tables, giving auditors a point-in-time rewind and giving DSPy the metrics to optimize its own routing and prompt choices.
  

The same guard-rails, then, that satisfy and delight regulators and humans experiencing the system, also feed the feedback loop that hones the agents’ reasoning—compliance, transparency, and relentless self-improvement, all drawn from the same telemetry well.

A Multi-Agent Architecture for Autonomous Data Warehousing

LangGraph gives the ensemble its backbone—a state-machine compiler that routes branches, retries, and roll-backs while preserving a single, shared state object (LangChain 2025) . Over that scaffolding runs DSPy, Stanford’s declarative, self-improving layer; instead of hand-tuned prompts we write ordinary Python that compiles into prompt logic (StanfordNLP 2024) . Within DSPy we implemented a Self-Discovery module after the framework described by Zhou and colleagues, so the system can decide—at run-time—when to reflect, when to spin out a tree of thought, and when to invoke a tool (Zhou et al. 2024) . Every one of those decisions is stamped into a LangGraph checkpoint and streamed into BigQuery, turning what would be ephemera into queryable facts (LangChain 2025b) .

The first worker is the Reflective SQL agent. It interrogates raw schemas the way an ethnographer maps kinship, stores each inference—episodic, semantic, procedural—in LangMem, and persists the lot to BigQuery for durable recall (LangChain 2025c) .

Next comes the DBT Data-Vault agent. Reading that ontology, it emits AutomateDV YAML that lets dbt spin up a full Data Vault 2.0 warehouse—hubs, links, satellites—without a white-board session (AutomateDV 2024; Khedekar 2024) .

The Analytics agent closes the loop. It queries the vault and paints the answers in Apache Superset dashboards or slips a chart into Slack, so a publican asking “Whisky versus vodka, last six months” sees the trend before the ice sweats (Superset 2025) .

LangSmith traces every prompt, token, and tool call, while a Unified Persistence Manager writes turn-level KPIs into clustered BigQuery tables (LangChain 2025d) . DSPy’s optimiser then drinks from the same tables, weighting tree depth, memory yield, latency, cost, and rapport to revise its own routing—guard-rails that sharpen, rather than blunt, the system’s reasoning.

Because all state lives in a regional BigQuery project, data remain inside jurisdictions that meet GDPR and similar statutes, and a compliance officer can rewind any run by filtering on timestamp alone (Google Cloud 2025; Galaxy 2025) . That same transparency also satisfies the explicability clause in the incoming EU AI Act, proving that ethics and regulation can be designed through, not patched on (Reuters 2025) .

Persistence and Telemetry: Why Store Agent Data in a Warehouse?

A cornerstone of NICER-DW is radical persistence. Every prompt, tool call, memory write and evaluator score is streamed into Google BigQuery rather than evaporating in RAM or a single-tenant vector store. LangGraph’s checkpoint API turns the whole state machine into versioned rows, so a swarm can resume from any branch without forfeiting context (LangGraph Docs 2025). That pipeline swaps the demo-friendly Postgres seen in most tutorials for columnar tables that scale to petabytes while keeping SQL as the lingua franca (Hsia 2025; Google Cloud BigQuery 2025).

Because nothing is thrown away, we gain provenance for free. LangSmith already captures each LLM interaction as a run-tree with latency, token counts and cost metadata; exporting those traces to the same warehouse lets auditors replay the exact chain of thought behind any decision, down to the moment the Reflective SQL worker inferred that CUST_ID links sales and users (LangSmith Docs 2025).

Treating telemetry as data also lets the agents study themselves. BigQuery’s SQL engine and BigQuery ML run directly over the log tables, revealing retry hot-spots, latency cliffs or falling coverage without shipping a byte elsewhere (Google Cloud Logging 2025). Vector search—now native in BigQuery—means the swarm can retrieve past prompts or reflections semantically, shortening the feedback loop between experience and memory (Google Cloud Vector Search 2024).

Scale and portability come baked in. BigQuery’s decoupled architecture elastically scans petabytes, yet its SQL dialect already federates into Snowflake, Synapse or an on-prem Oracle if residency rules demand it (Snowflake Federated 2023; Google Cloud BigQuery 2025). Swap the connection string and the checkpoint schema follows.

Security is woven, not stapled. BigQuery encrypts every byte at rest with AES-256 by default and supports customer-managed keys for the extra-cautious (BigQuery Encryption 2025). Fine-grained IAM, column tags and row-level security guard sensitive slices of data—whether they hold personal records or an agent’s speculative join (Google Cloud IAM 2025; Row-Level Security 2025). Region pinning meets GDPR; HIPAA and PCI attestations cover health and payments workloads out of the box (Google Cloud Compliance 2025).

In short, the warehouse is both product and mirror: the database our users query for insight is the same substrate the swarm consults to refine its own cognition. Observability, governance and optimisation collapse into one loop—an architecture that satisfies regulators and, more to the point, lets the agents learn faster because every move they make is queryable.

Global Compliance by Design

Compliance is never a retrospective patch in NICER-DW—it is the load-bearing wall. From the first design sketch we asked how a global bank, a London hospital and a Berlin start-up could all run the same swarm without breaching their own statutes of privacy and audit. The answer is threefold: keep data in the right place, keep proofs of every decision, and keep control planes that can say “no” faster than any agent can act.

Data never strays beyond agreed borders. BigQuery lets us pin every table to an EU or UK region, satisfying the GDPR’s strict localisation clauses and the UK Data Protection Act 2018 in a single stroke (Google Cloud Data-Residency 2025) (ICO 2018) (ICO UK GDPR 2025) . Agents see only the columns they need; retention jobs cull logs once statutory clocks expire, honouring GDPR’s data-minimisation principle.

Every step the swarm takes is indelible. LangGraph checkpoints plus LangSmith run-trees write a forensically complete journal—inputs, outputs, token counts, cost, latency—into the warehouse itself. Those records meet the EU AI Act’s call for “traceability and human oversight” by giving auditors, or sceptical managers, a scroll-back of the model’s reasoning, not a hand-wave (AuditBoard EU AI Act 2025) . If a recommendation edges into a sensitive zone, a human reviewer can replay the chain of thought before the business acts.

Sri Rang’s reference blueprint called for a policy engine, an audit ledger and a rollback valve; we instantiate each element. PermitRetriever in LangChain enforces per-agent permissions down to the tool call, so a mis-aligned worker cannot query a payroll table it has no business seeing (LangChain Permit 2025) . The Unified Persistence Manager stores every state delta with version IDs, letting us roll the system back to any millisecond with a single WHERE timestamp

Legal frameworks differ, but the controls converge. HIPAA demands encryption and audit for health data; BigQuery meets both under its covered-service attestation (Google Cloud HIPAA 2025) . PCI-DSS version 4.0 insists on secure processing of card information; Google’s own certification carries that burden for us (Google Cloud PCI DSS 2025) while our own logs prove processing integrity (UpGuard PCI Guide 2025) . ISO 27001’s Annex A lists logging and audit as mandatory controls; our checkpoint tables map one-to-one onto those clauses (ISMS.online 8.15 Logging 2022) . SOC 2 asks for evidence across the five Trust Service Criteria; BigQuery’s immutable tables and our row-level security satisfy security, availability, confidentiality and integrity in a single control surface (CSA SOC 2 Criteria 2023) .

Modularity keeps the swarm future-proof. If Brazil’s LGPD or India’s DPDP Act adds a disclosure field, we expose it in the catalog schema and the same governance SQL applies. Should a client insist on Snowflake or an on-prem Oracle, only the persistence adapter changes—the policy engine, audit SQL and rollback semantics travel intact (Snowflake Federated 2023) . Compliance, in other words, is not a moat we dig afterwards but the soil in which the whole system grows.

Ethical AI: From Capability Approach to Digital Humanism

Ethics, for us, is not a decorative “principle” stapled on after the prototype works; it is the criterion against which the system earns the right to exist; it’s raison d’etre. Two intellectual traditions shape that criterion.

Amartya Sen’s Capability Approach judges progress by the real freedoms people gain rather than by the goods they accumulate (Sen 1999). With the design of NICER-DW therefore we care less about latency records than about whether a pub-owner without a data team can spot a slump in whisky sales and act so as to pursue opportunities backed by reliable knowledge. By turning chat questions into SQL, and SQL into clear plots, the swarm converts expertise that once belonged to analysts into a capability everyone can wield. The warehouse is not a vault of numbers; it is a lever for agency.

Digital Humanism insists that technology should amplify, rather than attenuate, human agency (Vienna Manifesto 2019). Floridi frames that respect as “explicability”—the duty to let people understand and contest algorithmic conclusions (Floridi et al. 2018) . Our checkpoint logs satisfy the duty: any conclusion a bot draws is one SQL query and thus one sql agent invocation away from scrutiny (the agent allows you to ask in English). Human-in-the-loop controls—mandated by the EU AI Act’s human-oversight clause—can pause, annotate or veto a recommendation before it travels downstream (EU AI Act 2024) .

Yet ethical ambition falters without metrics. When a Reflective worker chooses which tables to scan, or the Superset worker auto-selects a colour scale, hidden defaults shape what the user sees. Left unchecked, such choices can skew judgement as surely as a rogue model can mis-score credit. NICER-DW therefore treats bias as a first-class KPI, mined from the same telemetry that feeds Insight-per-Dollar.

Every artefact—query text, result set, chart JSON, even the narrative caption an LLM writes—is snap-shotted to BigQuery. A nightly job casts those snapshots into IBM’s AI Fairness 360 schema and computes Statistical Parity Difference, Disparate-Impact Ratio, and Average-Odds Difference (Bellamy et al. 2018) . Thresholds follow common guidance—e.g., SPD above 0.10 or DIR outside 0.8-1.25 raises a yellow flag (Suresh & Guttag 2021; ECPG 2024) . Each score lands in ethical_bias_metrics, where a Bias-per-Dollar gauge sits beside Insight-per-Dollar, so ethical drift appears before it festers.

Why monitor a “plain” SQL stack so closely? Because bias hides in mundane corners. Representation bias appears when the SQL generator defaults to the last 30 days and silences historical context; measurement bias sneaks in when a “clean-up” drops negative values that correlate with low-income customers; aggregation bias flares when an average replaces a median on a log-skewed salary column; presentation bias blooms when a chart collapses minorities into an “Other” slice (McGregor et al. 2020) ose findings because carbon cost is a tax on everyone, making every human a stakeholder.

Finally, the whole cycle—measure, explain, mitigate, repeat—matches Google’s “human-centred development loop” for responsible AI (Google Research 2025) . Audit logs feed DSPy optimisation; optimisation improves insight yield; new logs check fairness, coverage, and energy again. Ethics here is not a sermon—it is an empirical subsystem, its metrics plotted beside revenue and latency.

Real-World Impact: Empowering Users with AI

Small firms are awash with ledgers, supplier feeds, and a trickle of social chatter, yet few have the time—or the analysts—to knit those streams into defensible decisions. NICER-DW closes that gap by letting owners question their data in English and receive an auditable, machine-generated answer composed from the same warehouse that stores the receipts.

A publican, for instance, can pick up her tablet and ask: “Show me whisky and vodka sales for the past six months.” The Analytics worker translates the request into SQL, the warehouse returns the numbers, and a Superset chart appears before the ice has melted. A glance shows whisky flat, vodka spiking in July; no menus, no pivot tables—just an evidence-grounded prompt–response (Superset Project 2025).

The same interface can chase questions that outstrip any accounting package. Suppose she now wonders whether to stock zero-proof cocktails. A forthcoming Social-Listening worker will stream public posts that mention “mocktails” and cognate terms, classify sentiment with a language model, and load rolling aggregates—volume, mood, geography—into BigQuery (IWSR 2025; Bellamy et al. 2018). Another prompt—“Compare mocktail sentiment with low-alcohol beer over the last quarter”—produces a paired time-series. If sentiment skews sharply positive, the system can nudge her: add two virgin cocktails before summer; younger customers are signalling demand. Because every intermediate query, sample, and model score is persisted, she can drill down to the exact forum posts that raised the flag—an enactment of Floridi’s “explicability” duty (Floridi et al. 2018).

Future workers extend the pattern. A causal-inference worker will link weather data, social sentiment, and sales, surfacing probabilistic explanations—heatwaves depress whisky, influencer buzz lifts mocktails—along with a confidence score and the trace that produced it (McGregor et al. 2020; Google Vector Search 2024). Because the Bayesian graph, its priors, and its fit diagnostics all live in the warehouse, an accountant or regulator can replicate the claim with a single SQL call.

This integration is what the Capability Approach celebrates: genuine freedom to act, not just data ownership (Sen 1999; Stanford SEP 2024). And because each worker’s defaults—sampling windows, colour scales, caption templates—are logged and nightly-scored with AI Fairness 360, any emergent bias becomes a KPI, tracked next to cost and latency (Suresh & Guttag 2021; Zhou et al. 2024). The pub owner, in short, gains a pocket analyst whose every assumption is both visible and negotiable—a tool that trades partisan rhetoric about “insight” for verifiable evidence she can act on.

Conclusion and Acknowledgments

The development of our multi-agent data warehouse system has been a journey in blending cutting-edge AI engineering with “old-fashioned” values of responsibility, transparency, and user-centric design. By orchestrating reflective learning agents with tools like LangGraph and DSPy, persisting their knowledge in an auditable BigQuery repository, and adhering to global compliance standards, we’ve aimed to build not just a clever system, but a trustworthy and inclusive one. We’re excited about the road ahead – integrating with more data platforms (Snowflake, Azure, etc.), adding new agents (the social listener and beyond), and continuously learning from real-world use.

Finally, a special thanks to the LangChain London community and the speakers at this week’s meetup. Events like these – with people generously sharing their knowledge and enthusiasm – are the fertile ground where ideas like ours take root and grow. We’re grateful to Marlene, Pablo, Sri, Nuno, and the organizers for creating such a joyful space for innovation . It’s clear that when passionate people come together (with perhaps a beverage in hand and plenty of good humor), technology moves forward in ways that truly serve people. Here’s to many more convivial meetups and collaborative breakthroughs on the path towards a more human-centered AI future!

Sources:

Anthropic (2024) Introducing the Model Context Protocol. 25 November. Available at: https://www.anthropic.com/news/model-context-protocol (Accessed: 23 July 2025).

Apache Software Foundation (2025) Apache Superset 4.1.3 Release Notes. 11 July. Available at: https://github.com/apache/superset/releases/tag/4.1.3 (Accessed: 23 July 2025).

AuditBoard EU AI Act (2025) Navigating New Regulations for AI in the EU. Available at: https://auditboard.com (Accessed: 23 July 2025).

AutomateDV Project (2024) AutomateDV Documentation. Available at: https://automate-dv.readthedocs.io (Accessed: 23 July 2025).

Bellamy, R.K. et al. (2018) ‘AI Fairness 360: A Comprehensive Toolkit for Fairness in AI’, IBM Research Technical Report. Available at: https://arxiv.org/abs/1810.01943 (Accessed: 23 July 2025).

Campos, N. (2025a) ‘LangGraph 0.3 release: Pre-built agents’, LangChain Blog, 27 February. Available at: https://blog.langchain.com/langgraph-0-3-release-prebuilt-agents/ (Accessed: 23 July 2025).

Campos, N. (2025b) ‘Uniting agentic and frontend code bases feels natural’, LinkedIn post, 4 March. Available at: https://www.linkedin.com/posts/nuno-f-campos (Accessed: 23 July 2025).

Cartesia AI (2024) ‘State of Voice AI 2024’. 12 December. Available at: https://cartesia.ai/blog/state-of-voice-ai-2024 (Accessed: 23 July 2025).

CSA (Cloud Security Alliance) (2023) The Five SOC 2 Trust-Services Criteria Explained. Available at: https://cloudsecurityalliance.org (Accessed: 23 July 2025).

ECPG (European Committee on Payment Guidelines) (2024) Guidelines on Statistical Parity.

European Parliament (2024) Artificial Intelligence Act – Provisional Text. Brussels: EU Publications Office.

Floridi, L. et al. (2018) ‘AI4People—An Ethical Framework for a Good AI Society’, Minds and Machines, 28(4), pp. 689-707.

Google Cloud (2025a) BigQuery Emerges as Autonomous Data-to-AI Platform. Available at: https://cloud.google.com/blog/products/data-analytics/bigquery-emerges-as-autonomous-data-to-ai-platform (Accessed: 23 July 2025).

Google Cloud (2025b) Platform Services Supporting Data Residency. Available at: https://cloud.google.com/security-command-center/docs/data-residency-support (Accessed: 23 July 2025).

Google Cloud BigQuery (2025) ‘AI Data Platform Overview’. Available at: https://cloud.google.com/bigquery/docs (Accessed: 23 July 2025).

Google Cloud Compliance (2025) BigQuery Compliance Resource Centre. Available at: https://cloud.google.com/bigquery/docs/compliance (Accessed: 23 July 2025).

Google Cloud Encryption (2025) Encryption at Rest. Available at: https://cloud.google.com/docs/security/encryption (Accessed: 23 July 2025).

Google Cloud HIPAA (2025) HIPAA Compliance on Google Cloud. Available at: https://cloud.google.com/security/compliance/hipaa (Accessed: 23 July 2025).

Google Cloud IAM (2025) ‘Column-level Access Control’. Available at: https://cloud.google.com/bigquery/docs/column-level-security-intro (Accessed: 23 July 2025).

Google Cloud Logging (2025) ‘Query and Analyse Logs with Log Analytics’. Available at: https://cloud.google.com/logging/docs/query-analytics (Accessed: 23 July 2025).

Google Cloud PCI DSS (2025) PCI DSS Compliance Resource Center. Available at: https://cloud.google.com/security/compliance/pci-dss (Accessed: 23 July 2025).

Google Cloud Vector Search (2024) ‘Vector Search for Log Analysis’. Google Cloud Blog, 9 October. Available at: https://cloud.google.com/blog/products/ai-machine-learning/vector-search-for-log-analysis (Accessed: 23 July 2025).

Google Research (2025) Human-Centred Development Cycles for Responsible AI. Mountain View: Google AI.

Hsia, W. (2025) ‘Easy analytics using LangGraph checkpoints with BigQuery’, Google Cloud Community, 8 May. Available at: https://cloudcommunity.dev (Accessed: 23 July 2025).

ICO (Information Commissioner’s Office) (2018) Data Protection Act 2018: Overview. Wilmslow: ICO.

ICO UK GDPR (2025) UK GDPR Guidance and Resources. Available at: https://ico.org.uk (Accessed: 23 July 2025).

IBM (2025) ‘What is Model Context Protocol (MCP)?’. Available at: https://www.ibm.com/think/topics/model-context-protocol (Accessed: 23 July 2025).

ISMS.online (2022) ‘ISO 27001:2022 Annex A — 8.15 Logging’. Available at: https://www.isms.online (Accessed: 23 July 2025).

LangChain (2025a) ‘LangGraph BigTool for providing agents access to a large number of tools’, LangChain Changelog, 19 March. Available at: https://changelog.langchain.com (Accessed: 23 July 2025).

LangChain (2025b) ‘LangMem SDK for Agent Long-Term Memory’, LangChain Blog, 18 February. Available at: https://blog.langchain.com/langmem-sdk-launch/ (Accessed: 23 July 2025).

LangChain (2025c) Tracing with LangGraph (Python and JS/TS). LangSmith Documentation. Available at: https://docs.smith.langchain.com/observability/how_to_guides/trace_with_langgraph (Accessed: 23 July 2025).

LangChain (2025d) Observability Quick Start. LangSmith Documentation. Available at: https://docs.smith.langchain.com/observability (Accessed: 23 July 2025).

LangChain (2025e) How to Query Traces and Runs. LangSmith Docs. Available at: https://docs.smith.langchain.com (Accessed: 23 July 2025).

LangChain London Meetup (2025) Agenda and Speakers.

LangChain PermitRetriever (2025) Documentation.

LangGraph Docs (2025) ‘Persistence and Checkpoints’. LangChain Documentation. Available at: https://docs.langgraph.ai (Accessed: 23 July 2025).

McGregor, A. et al. (2020) ‘Visualising Bias’, in Proceedings of CHI 2020. New York: ACM.

Mhangami, M. (2025) Getting Started with Model Context Protocol (MCP). Microsoft Reactor event. Available at: https://developer.microsoft.com/en-us/reactor/events/25883 (Accessed: 23 July 2025).

Microsoft Reactor (2025) ‘Microsoft Reactor London | Meetup’. Available at: https://www.meetup.com/microsoft-reactor-london/ (Accessed: 23 July 2025).

Patel, K. (2025) ‘From Prototype to Production: Why LangChain Revolutionises Enterprise AI’, LinkedIn article. Available at: https://www.linkedin.com/pulse/from-prototype-production-why-langchain-enterprise-ai-patel (Accessed: 23 July 2025).

Reuters (2025) ‘AI models with systemic risks given pointers on how to comply with EU AI rules’. Reuters, 18 July. Available at: https://www.reuters.com/world/europe/ai-models-systemic-risks-comply-eu-ai-rules-2025-07-18 (Accessed: 23 July 2025).

Row-Level Security (2025) ‘Introduction to Row-Level Security’. Google Cloud Documentation. Available at: https://cloud.google.com/bigquery/docs/row-level-security-intro (Accessed: 23 July 2025).

Scott, K. (2025) ‘Microsoft CTO Kevin Scott on how AI can save the web, not destroy it’, The Verge, 21 May. Available at: https://www.theverge.com/decoder-podcast-with-nilay-patel/669409 (Accessed: 23 July 2025).

Sen, A. (1999) Development as Freedom. Oxford: Oxford University Press.

Snowflake Federated (2023) ‘Executing Federated Queries in Snowflake Using Google BigQuery’. Medium, 14 September. Available at: https://medium.com (Accessed: 23 July 2025).

Stanford Encyclopedia of Philosophy (2024) ‘Capability Approach’. Available at: https://plato.stanford.edu/entries/capability-approach (Accessed: 23 July 2025).

Stanford HAI (2024) DSPy: Compiling Declarative Language-Model Calls into State-of-the-Art Pipelines. Available at: https://hai.stanford.edu/research/dspy-compiling-declarative-language-model-calls-into-state-of-the-art-pipelines (Accessed: 23 July 2025).

Suresh, H. and Guttag, J. (2021) ‘Sources of Harm in the Machine-Learning Life Cycle’, in Proceedings of ACM FAccT 2021.

UNESCO-UCL (2025) Energy Footprints of Large Language Models: Measurement and Mitigation. Paris: UNESCO.

UpGuard (2025) ‘How to Comply with PCI DSS 4.0.1’. Available at: https://www.upguard.com (Accessed: 23 July 2025).

Vienna Manifesto (2019) Manifesto on Digital Humanism. Vienna University of Technology.

Wei Hsia, LangChain AI (2025) ‘Easy Analytics Using LangGraph Checkpoints with BigQuery and Cloud SQL’, Google Cloud Community, 1 June. Available at: https://medium.com/google-cloud/easy-analytics-using-langgraph-checkpoints (Accessed: 23 July 2025).

Zhou, P. et al. (2024) ‘Self-Discover: Large Language Models Self-Compose Reasoning Structures’, arXiv pre-print arXiv:2402.03620. Available at: https://arxiv.org/abs/2402.03620 (Accessed: 23 July 2025).