concrete examples
What you'd actually build.
Three real operational workflows, end to end: inputs, run trace, artifacts, receipt. Same runtime primitives across all three.
Security · Compliance · CISO
soc2-autonomous-agent-mesh
Flagship 4-agent mesh: evidence collector → policy reviewer → vulnerability scanner → compliance approver. Quarterly SOC2 evidence pack with replay + receipts for the auditor.
outcome → 3 minutes + one scope approval. Auditor gets the evidence pack and a replay handle to re-run the mesh deterministically. The receipt is the audit artifact.
inputs
- policies/14 PDFs · v2026Q1
- access-reviews/q1-2026.csv1,842 rows
- vendors/27 SOC reports
- tickets/security-incidents.jsonJira export
- scans/Tenable + Snyk feeds · 89 services
- runbooks/47 approved playbooks
run trace · live SSE
- 00:00:02planner-agent: dispatch 4 specialists · scoped grants per role
- 00:00:03evidence-collector: grant policies/ + access-reviews/ + vendors/ + tickets/ (read)
- 00:00:04policy-reviewer: grant policies/ (read) · CC1–CC9 control matrix
- 00:00:05vulnerability-scanner: grant scans/ (read) · CVE + service inventory
- 00:00:14evidence-collector: extracted 312 evidence items · cross-referenced 1,842 access rows
- 00:00:41policy-reviewer: 47 controls scored · 44 covered · 3 partial
- 00:00:58
- 00:01:02vulnerability-scanner: 12 high-sev open · 4 inside SLA · 8 overdue
- 00:01:18evidence-collector: request_scope("vendors/offboarding/*", reason="CC6.7")
- 00:01:52approver @aisha approved scope expansion · grant rotated (60-min)
- 00:02:33handoff → compliance-approver: 3 specialist reports → control matrix
- 00:02:48
- 00:03:11artifact: evidence-pack/q1-2026.zip (217MB · 489 files)
- 00:03:13receipt sealed · replay handle attached · eval 0.94 · ready for CISO
artifacts
- evidence-pack/q1-2026.zip217MB · 489 files · auditor-ready
- evidence-pack/control-matrix.xlsxCC1–CC9 · 47 controls · per-agent attestation
- evidence-pack/exceptions.md3 items · CISO sign-off queued
- evidence-pack/replay-handle.urldeterministic re-run for auditor verification
- evidence-pack/mesh-graph.svgagent topology · handoff lineage
receipt.json
- meshsoc2-autonomous-agent-mesh@v2.0.0
- agents4 (collector + reviewer + scanner + approver) + planner
- callercompliance-lead@acme
- handoffs4 (planner → 3 specialists → approver)
- grants4 scoped + 1 ephemeral (rotated for CC6.7)
- approvals1 (scope expansion · auditable)
- files2,847 read · 489 written
- elapsed3m 13s
- eval0.94 · CISO review queued
Sales Engineering · Revenue
rfp-responder
Read inbound RFPs, pull approved collateral, draft a cited response pack with reviewer checkpoints.
outcome → cited draft + chart + flagged questions in under 2 minutes. SE reviews 1 question, not 47.
inputs
- rfp/2026-globex.pdf47 questions · 128 pages
- collateral/security-faq.pdfapproved · v2026Q2
- collateral/case-studies/12 PDFs
- pricing/rate-card.jsonapproved tier
run trace · live SSE
- 00:00:01grant accepted · workspace/rfp/*.pdf + collateral/* (read-only)
- 00:00:08extracted 47 questions from globex.pdf
- 00:00:24matched 41 questions to approved collateral · 6 unmapped
- 00:00:38
- 00:00:39flagged Q19 for human · continuing 46 questions
- 00:01:47draft assembled · 46 cited answers · 1 flagged
- 00:01:48handoff → chart-agent ("render pricing comparison")
- 00:01:55artifact: rfp/draft/globex-response.docx (87 pages · 312 cites)
artifacts
- rfp/draft/globex-response.docx87 pages · 312 source cites
- rfp/draft/pricing-comparison.pngchart-agent handoff · 1.2MB
- rfp/draft/open-questions.md1 flagged · routed to SE lead
receipt.json
- agentrfp-responder@v2.1.0
- callerplanner-agent@v0.9
- handoffs1 (chart-agent)
- files16 read · 3 written
- elapsed1m 55s
- evalauto: 0.89 · review queued
Finance · Month-end close
finance-recon
Compare invoices against contracts, usage exports, and payments. Produce exception report + auditor-ready trace.
outcome → 3 minutes + 3 escalations. Trace links every exception back to its source files.
inputs
- invoices/2026-04/1,247 invoices · 89MB
- contracts/master.csv342 active MSAs
- usage/april.parquet27M events
- payments/ach-2026-04.csvTreasury export
run trace · live SSE
- 00:00:03grant accepted · workspace/{invoices,contracts,usage,payments}/2026-04/*
- 00:00:21normalized 1,247 invoices to canonical schema
- 00:01:08joined invoices × contracts on customer_id (1,201 matched)
- 00:01:14
- 00:02:33usage variance computed · 12 customers >5% delta
- 00:03:02
- 00:03:18artifact: reports/2026-04-exceptions.xlsx (61 rows)
- 00:03:20receipt sealed · linked to GL period 2026-04
artifacts
- reports/2026-04-exceptions.xlsx61 rows · 3 critical · 58 review
- reports/2026-04-summary.pdfexec-ready · 4 pages
- reports/2026-04-trace.jsonl27,847 events · auditor-grade
receipt.json
- agentfinance-recon@v1.4.2
- callercontroller@acme
- grants1
- files1,250 read · 3 written
- elapsed3m 20s
- eval0.91 · reviewed @max
Corp Dev · Outside Counsel
legal-due-diligence
Due-diligence swarm on a target data room: extract contracts, classify risk, verify citations, summarize for partners.
outcome → Standard DD timeline: 3 weeks of associates. With the swarm: 7 minutes + partner sign-off. Every flagged risk has a receipt link to the source clause.
inputs
- dataroom/contracts/1,847 PDFs · 6.2GB
- dataroom/cap-table.xlsx12 series · 287 holders
- dataroom/ip-assignments/412 agreements
- dataroom/litigation/23 matters · 8 active
run trace · live SSE
- 00:00:04grant accepted · workspace/dataroom/* (read-only, partner-audience-bound)
- 00:00:18contract-extractor: 1,847 PDFs → structured clauses (442k clauses)
- 00:02:11handoff → risk-classifier: change-of-control, IP assignment, exclusivity, MFN
- 00:04:47
- 00:05:02handoff → citation-verifier: 1,204 cross-references checked
- 00:06:13
- 00:06:40handoff → exec-summarizer: partner-ready risk memo
- 00:07:22artifact: dd-memo/target-co-2026Q2.pdf (47 pages · 312 cites · 71 risks)
- 00:07:24receipt sealed · eval 0.93 · partner sign-off queued
artifacts
- dd-memo/target-co-2026Q2.pdf47 pages · 312 cites
- dd-memo/risk-register.xlsx71 risks · severity ranked
- dd-memo/clause-extracts.parquet442k clauses · queryable
- dd-memo/open-questions.md9 items · routed to outside counsel
receipt.json
- agentlegal-due-diligence@v0.8.1
- callerpartner@firm
- handoffs3 (extractor → classifier → verifier → summarizer)
- grants1 (audience-bound to partner)
- files2,694 read · 4 written
- elapsed7m 24s
- eval0.93 · partner review queued
R&D · Translational Science
pharma-research
Read the literature, design the next experiment, validate the protocol, route compliance. Provenance for every claim.
outcome → From literature to a compliance-clean experiment design in 8 minutes. Every claim links back to the paper, every protocol clause links back to the template — provenance is the artifact.
inputs
- literature/oncology-q2/2,140 papers · PubMed export
- internal/assay-results/417 runs · 12 cohorts
- protocols/templates/GCP / GLP / IRB
- compliance/policies/21 CFR Part 11 · HIPAA
run trace · live SSE
- 00:00:05grant accepted · workspace/literature + assay-results (read), protocols/templates (read)
- 00:00:42paper-analyzer: 2,140 papers → 1,847 indexed (293 paywalled, skipped)
- 00:03:18joined external findings × internal assay-results on target / mechanism
- 00:04:51
- 00:04:55handoff → experiment-planner: dose-response design v3
- 00:06:34handoff → protocol-validator: GLP + IRB template check
- 00:07:11
- 00:07:48handoff → compliance-reviewer: 21 CFR Part 11 audit trail check
- 00:08:22artifact: exp-2026-Q2-017.pdf (32 pages · 412 cites · provenance graph)
- 00:08:24receipt sealed · ready for PI review
artifacts
- experiments/exp-2026-Q2-017.pdf32 pages · 412 cites
- experiments/provenance.jsonldlinked data · cite-to-claim graph
- experiments/dose-response-design.xlsx8 arms · 240 samples
- experiments/compliance-attestation.md21 CFR Part 11 · GLP · IRB
receipt.json
- agentpharma-research@v0.5.0
- callerpi@lab
- handoffs4 (analyzer → planner → validator → reviewer)
- grants2 (literature read-only, protocols read-only)
- files2,558 read · 4 written
- elapsed8m 24s
- eval0.88 · PI review queued
Security Operations · Detection & Response
secops-mesh
SIEM signal → CVE triage → remediation plan → human approval gate. Authority delegation, runtime isolation, full replay.
outcome → patched prod with a human in the loop, in under 10 minutes. Authority for the patch agent expires at the 60-min mark — no standing deploy keys, no ambient blast radius.
inputs
- siem/alerts/2026-05-28/1,247 alerts · 12 high-sev
- cve/feed-latest.jsonMITRE + GHSA · 89 new
- inventory/services.parquet317 services · 1,840 deps
- runbooks/47 approved playbooks
run trace · live SSE
- 00:00:02grant accepted · siem/alerts (read), inventory (read), runbooks (read)
- 00:00:09siem-analyzer: triaged 1,247 alerts → 12 high-sev · 3 likely-true-positive
- 00:00:34handoff → cve-triager: matched alerts to CVE-2026-4471 (golang/net)
- 00:01:18
- 00:01:22handoff → remediation-planner: ordered patch plan · canary first
- 00:02:04plan: bump golang/net to v0.42.1 · 4 services rolling restart · 19 batch
- 00:02:07
- 00:03:11approver @oncall approved · scoped delegation · 60-min window
- 00:03:14patch agent running in microVM · 4 services patched · health green
- 00:09:42artifact: incident-2026-05-28.md · receipt + replay link
artifacts
- incidents/2026-05-28-cve-2026-4471.mdpost-mortem · auditor-ready
- incidents/patch-plan.yamlordered · approved · executed
- incidents/replay-link.urldeterministic re-run handle
- incidents/affected-services.csv23 services · before/after
receipt.json
- agentsecops-mesh@v0.7.3
- calleroncall@acme
- handoffs3 (analyzer → triager → planner → patch-agent)
- grants1 scoped + 1 ephemeral (60-min deploy window)
- approvals1 (oncall · auditable)
- files1,896 read · 27 written
- elapsed9m 42s
- eval0.92 · post-mortem queued
Data Platform · Analytics Engineering
etl-autonomous
Ingest, transform, validate, load — without a hand-written DAG. Schema drift detected, contracts enforced, lineage captured.
outcome → No DAG, no Airflow, no on-call page at 3am. Schema drift becomes a filed issue, not a 200-line null check. Lineage is automatic — every dashboard can trace upstream to the source file.
inputs
- raw/events/2026-05-28/12 GB · 84 parquet files
- schemas/contracts.yaml47 active contracts
- warehouse/snowflake.credsscoped role · DBT writer
- policies/pii-redaction.yamlGDPR + CCPA
run trace · live SSE
- 00:00:03grant accepted · raw/events (read), warehouse (write:fact_events_v3)
- 00:00:21ingest-agent: 84 files validated against contract events@v3
- 00:00:48
- 00:00:52handoff → transform-agent: backfill nulls + flag contract update
- 00:01:34pii-redactor: hashed user_id · stripped raw_ip · GDPR-clean
- 00:02:11handoff → validate-agent: row counts ±0.1%, dupes 0, nulls within bounds
- 00:02:38handoff → warehouse-loader: COPY INTO fact_events_v3 · transaction-wrapped
- 00:03:42loaded · 142M rows · 12.4 GB · lineage attached
- 00:03:44artifact: lineage-graph + contract-drift-issue + load receipt
artifacts
- lineage/fact_events_v3-2026-05-28.jsonldOpenLineage · upstream + downstream
- issues/contract-drift-events-v3.mdfiled · routed to data PM
- warehouse/load-report.jsonrows / bytes / partitions / time
- audit/pii-redaction-attestation.txtGDPR + CCPA signed
receipt.json
- agentetl-autonomous@v0.4.2
- callerdata-platform@acme
- handoffs4 (ingest → transform → validate → loader)
- grants1 read + 1 scoped write (table-bound)
- files84 read · 142M rows written
- elapsed3m 44s
- eval0.95 · contract drift auto-filed
all three share
Same primitives. Different work.
Identity. Scoped grants. Sandboxed runtime. Workspace files. Approvals. Receipts. Replay. Each workflow above is @a2a.tool + runtime — no bespoke deployment pipeline, no shared credentials, no audit gap.