agentflow-runtime 1.5.0

Real-time data platform serving context to AI agents

AgentFlow

Real-time data platform for AI agents. Live entity lookups, typed contracts, dual-language SDKs, and release-gated delivery.

Why this exists

Most agent demos work until they have to answer from live business state. Support, ops, and merch workflows need current orders, metrics, and health signals while the conversation is happening, not a stale warehouse snapshot and not a pile of one-off service adapters.

AgentFlow turns that problem into one serving boundary:

• streaming ingestion for operational events
• a semantic layer that exposes entities, metrics, and query endpoints
• typed contracts so SDKs and callers know what shape to expect
• Python and TypeScript clients that speak the same API surface

Highlights

• Release line through v1.4.0 on PyPI (agentflow-runtime, agentflow-client) and npm (@yuliaedomskikh/agentflow-client), published via OIDC Trusted Publishers with SLSA provenance attestations on every artifact. Live registry table + re-verify recipe: docs/dv2-multi-branch/RELEASE_STATUS.md
• 561 unit tests and the 842-test Windows no-Docker suite green locally; CI runs 12 required status checks (lint, schema-check, test-unit, test-integration, helm-schema-live, perf-check, terraform-validate, bandit, safety, npm-audit, trivy, contract). Branch protection requires every one of them
• Sub-second entity lookups: entity p50 38-55 ms, entity p99 167 ms on local hardware (–82% from the 2026-04-23 baseline after the PII masker + tenant qualification cache wins). CI runner thresholds are documented separately in docs/perf/ci-hardware-gap-2026-05-24.md
• Dual SDK parity for Python (agentflow-client) and TypeScript (@yuliaedomskikh/agentflow-client), including retry policies, circuit breakers, batching, pagination, contract pinning, idempotency keys, and as_of historical reads
• Two CDC paths: production-grade Debezium + Kafka Connect (Helm chart hardened with NetworkPolicy + PDB + securityContext, schema-validated on every deploy), and a ClickHouse MaterializedPostgreSQL per-branch fan-out for the DV2 demo cluster
• Security hardening in the hot path: tenant isolation across every read surface, parameterized queries, sqlglot AST validation for NL-to-SQL, fail-closed auth middleware, plaintext secret scrubbing, and a Bandit baseline gate for new findings only
• On-call playbooks for production incidents in docs/runbooks/: symptom-keyed runbooks for API 5xx spikes, auth fail-closed regressions, CDC lag, Load Test gate failures, and PyPI/npm release rollback
• DV2 demo triptych: voice-narrated terminal cast (demo_voiced.mp4, 92s) + web-UI screencast covering Argo Workflows and MinIO (demo_webui.mp4, 60s) + dbt docs lineage walk-through (demo_dbt_docs.mp4, 55s)

Quick start

Upgrading from v1.0.x? See the v1.1 migration guide before installing.

Prerequisites:

• Python 3.11+
• make
• Docker Compose (make demo starts Redis)

PowerShell 7+:

git clone https://github.com/brownjuly2003-code/agentflow.git
cd agentflow
. .\scripts\setup.ps1
make demo

macOS / Linux:

git clone https://github.com/brownjuly2003-code/agentflow.git
cd agentflow
source ./scripts/setup.sh
make demo

make demo seeds local data, starts Redis, and serves the API on http://localhost:8000. Swagger UI is available at http://localhost:8000/docs.

Try it:

curl http://localhost:8000/v1/entity/order/ORD-20260404-1001

curl -X POST http://localhost:8000/v1/query \
  -H "Content-Type: application/json" \
  -d '{"question":"Show me top 3 products"}'

Local demo runs without API-key enforcement unless you explicitly configure AGENTFLOW_API_KEYS_FILE.

Architecture

Event sources -> Kafka -> Flink -> Iceberg ----\
                                                -> Semantic layer -> FastAPI -> Agent / SDK
Local demo   -> local_pipeline -> DuckDB ------/

Stack:

• Ingestion: Kafka producers, Debezium/Kafka Connect CDC, and a local synthetic pipeline
• Processing: Flink plus validation and enrichment stages
• Storage: Iceberg for production-shaped tables, DuckDB for the local serving path
• Serving: FastAPI, contract registry, lineage, search, and operational endpoints
• Orchestration: Dagster
• IaC: Terraform, Helm, Docker Compose, and a Fly.io demo config

See docs/architecture.md for the detailed design, trade-offs, and deployment topologies.

CDC source capture is standardized on Debezium/Kafka Connect; downstream consumers use the canonical AgentFlow CDC contract defined in ADR 0005.

What's inside

Area	Files
API core	src/serving/api/
Semantic layer	src/serving/semantic_layer/
Python SDK	sdk/agentflow/
TypeScript SDK	sdk-ts/src/
Agent integrations	integrations/agentflow_integrations/ (LangChain, LlamaIndex, CrewAI, MCP)
Flink jobs	src/processing/flink_jobs/
Test suites	tests/
Planning trail	docs/plans/
Public site	site/
IaC	infrastructure/terraform/, infrastructure/dv2/, helm/, k8s/
DV2.0 warehouse	warehouse/agentflow/dv2/ (hubs / links / satellites + X5 loader)

Documentation

• Session Handoff - start here when picking up the project cold — current HEAD, open Dependabot PRs, externally user-gated items, recent lessons (Contract Tests path filter, Dependabot cascade conflicts)
• Interactive Technical Walkthrough - MkDocs Material guide with Mermaid architecture, API, SDK, deployment, observability, and troubleshooting pages
• Architecture - system context, data flow, failure modes
• Operational Runbook - local stack, CDC capture, and maintenance commands
• On-Call Runbooks - production-incident playbooks (API 5xx, auth break, CDC lag, Load Test regression, release rollback)
• API Reference - endpoint-by-endpoint examples for curl, Python, and TypeScript
• Security Audit - threat model, controls, and evidence
• Competitive Analysis - positioning and trade-offs
• CDC Deployment Plan - Debezium/Kafka Connect rollout decisions and implementation trail
• Glossary - interview-ready explanations of the core technical terms
• Release Readiness - checked release evidence through v1.1.0 and current external-gate handoffs
• Customer Discovery Tracker - synthetic/modelled PMF planning surface and real-evidence gap
• Pricing Validation Plan - pricing/WTP questions, evidence gates, and current real-evidence gap
• AWS OIDC Setup - archived optional Terraform apply path; currently out of scope because there is no AWS budget or foreign-card/payment path
• Immutable Retention Evidence Handoff - external audit-retention evidence checklist
• Production CDC Source Onboarding - production CDC decision record and no-go gates
• External Gate Evidence Intake Checklist - required owner evidence before blocked external gates can close
• Public Production-Hardware Benchmark Plan - benchmark evidence plan for c8g.4xlarge+
• External Pen-Test Attestation Handoff - third-party pen-test evidence checklist
• Audit History - baseline-to-release remediation trail
• Publication Checklist - final GitHub publishing checklist
• Fly.io Demo Deploy - minimal hosted demo instructions
• DV2.0 Multi-Branch Extension - Data Vault 2.0 model for mid-market e-com with 5 locations / 3 jurisdictions; schema, end-to-end flow, live demo evidence, k8s manifests
• Contributing - development and PR expectations
• Changelog - project release notes

Development

# verified release slice
python -m pytest tests/unit tests/integration tests/sdk -q

# benchmark and regression gate
python scripts/run_benchmark.py
python scripts/check_performance.py --baseline docs/benchmark-baseline.json --current .artifacts/load/results.json --max-regress 20

# benchmark trend: [.github/perf-history.json](.github/perf-history.json) is appended on every main push;
# render the history locally with `make perf-plot` (writes docs/perf/history.html).

# contracts and security
python scripts/generate_contracts.py --check
bandit -r src sdk --ini .bandit --severity-level medium -f json -o .tmp/bandit-current.json
python scripts/bandit_diff.py .bandit-baseline.json .tmp/bandit-current.json

Status

v1.4.0 is the current release line (PyPI agentflow-runtime / agentflow-client, npm @yuliaedomskikh/agentflow-client, all three published 2026-05-24T21:05Z via OIDC Trusted Publishers with SLSA provenance attestations). Live registry table and re-verify recipe live in docs/dv2-multi-branch/RELEASE_STATUS.md.

The v1.1.0 → v1.4.0 arc landed in four increments on top of the 2026-04-27 audit closure sprint:

• v1.1.0 — audit closure: tenant isolation across every read surface, SQL guard centralization on sqlglot, entity allowlist enforcement, fail-closed auth, secrets rotated, Helm hardening, npm audit clean, vulnerable dep bumps, OpenAPI drift gate, 12 required status checks, Python SDK alignment with server v1 contract (F1–F10).
• v1.2.0 — DV2 multi-branch warehouse merged to main: 38 DV2.0 tables (8 hubs / 8 links / 22+ satellites), Argo Workflows dv2-refresh template, dbt project with 3 mart models + 12 tests, per-branch CDC fan-out via ClickHouse MaterializedPostgreSQL, and the first voice-narrated terminal cast demo.
• v1.3.0 — helm/kafka-connect chart hardening matched to helm/agentflow (NetworkPolicy + PDB + pod/container securityContext
  • /tmp emptyDir, all schema-required and off-by-default), Helm live validation parametrized across both charts, A03 CI hardware-gap acceptance with Load Test gates raised to 1.3x baseline, and the DV2 demo triptych completed (terminal + web-UI + dbt docs screencasts).
• v1.4.0 — maintenance release: top-level handoff and release docs, on-call runbooks, SECURITY.md, issue/PR templates, contract/DORA CI hardening, Dependabot/editorconfig repo hygiene, type-stub adoption, and the Tier A dependency wave (mypy, Terraform AWS provider, TypeScript, GitHub Actions, Vitest). No runtime API changes from v1.3.0.

CI on main is fully green across all 12 required checks. Local hardware sustains entity p99 167 ms (the local SLO target); CI runner thresholds are intentionally divergent and documented in docs/perf/ci-hardware-gap-2026-05-24.md.

Remaining external gates require inputs outside this repository. AWS is not one of them for the current plan: Terraform apply is intentionally out of scope because there is no AWS budget or foreign-card/payment path, and the DV2/X5 demo uses the documented HF Datasets/Backblaze-compatible cold-tier path for derived/anonymized parquet.

• Production CDC source onboarding (hostnames, credentials, owners, private network path) — runbook ready in docs/operations/cdc-production-onboarding.md.
• Real PMF / pricing evidence and a public benchmark on production-grade hardware (c8g.4xlarge+).
• External pen-test attestation.
• Optional: paid larger GHA runner or self-hosted runner if the CI hardware-gap thresholds need to be tightened.

The project-local Pi skill at .pi/skills/external-gate-evidence-intake and the External Gate Evidence Intake Checklist codify what evidence must arrive before any of those gates close.

Screenshots

Admin UI	API docs

Landing page	Benchmark run

Capture notes and publish-time checks are listed in docs/publication-checklist.md.

License

MIT. See LICENSE.

Credits

Built as a data-engineering reference project. Initial release cycle 2026-04-10 → 2026-04-20, post-audit hardening and DV2 extension through 2026-05-25 (v1.4.0). Full implementation trail preserved in docs/plans/, docs/codex-tasks/, and docs/lessons/.