Three institutional agents. One durable investigation identity. No raw IAM binding crosses the model or human-notification boundary.
Technical report (10 November 2026): Bastion: Execution Boundaries, Control Evolution, and Bounded Evaluation of a Governed Agent Fleet, DOI 10.5281/zenodo.24120596. Not peer reviewed.
Status: the fleet ran on Google Cloud from August to November 2026 and was retired on 9 November 2026. This repository is the archived source, evidence and report. It began as an entry to the All Things Agentic Hackathon 2026 (Fortified Enterprise Fleet).
Between August and November 2026, five dated failures showed the same gap: a local signal did not establish that the intended work had run across an agent boundary. The signals were a completed status, a worker-local state key, a state-only policy decision, an error-free remote reply and a closed response stream. Three of the failures escaped a test suite with 100% configured statement and branch coverage, because the tests exercised an in-process topology in which state is shared.
The repairs made policy execution, handoffs and downstream completion explicitly checkable. The technical report documents the failures, the repairs and a bounded evaluation on one project and one workload. It is not a reliability rate or a general security guarantee.
Bastion reviews the real IAM policy of the GCP project that runs it, using read-only Cloud Asset Inventory. Deterministic code flags Owner/Editor grants, scores them, applies human exceptions and routes reviews. Gemini writes a bounded rationale; it does not decide risk or delivery. Raw members, roles, resources and bindings stay inside the Auditor's deterministic tool. The minimised schema carries opaque IDs, departments, categories and scores, not an anonymity or general zero-leakage guarantee.
Approved event → Pub/Sub/Eventarc → IAM-protected durable dispatcher → managed Agent Runtime → Gateway/Registry/IAP → Access Auditor → deterministic policy step and gate → model-free Escalation Agent → completion gate.
There are three institutional agents; the policy steps and gates are internal deterministic steps. The dispatcher has no local agent graph or direct worker path. A2A workers validate the deployment-owned origin credential; the findings API uses Cloud Run IAM. There is no public agent demo endpoint. The separate evidence console serves static sanitised records and makes no request to the private fleet.
The claim matrix bounds each claim; the evidence index records measurements and their limits.
- The dated August captures establish the governed route, identity separation, minimised real-IAM findings, screening, audit and trace observations. The inventory records 21 APIs and 39 resources; it is not a current health check.
- The 27 October pilot shaped the pre-defined November cohort N. N ran on the unchanged August deployment and retained its false completion alongside a successful batch. It is not a reliability rate.
- On 5 November, both deployment surfaces were updated. The first F pilot failed closed on an Escalation routing parser (E24). Separately corrected recovery passed five normal cases, a direct confirmation control and two replays; the refusal join was verified read-only after a collector false positive (E25). The original stop is retained.
- The 6 November daily billing check passed for currently reported usage; exact per-chunk currency attribution and late postings remain bounded. Current balances and reservations stay private.
- Version provenance separates source pins, image assertions and unreadable Runtime installed packages. v0.9.2 (11 November) is the current source release and was never deployed; v0.9.1 (7 November, ADK 2.14.0) is the final deployment the report evaluates (evidence 26–29); v0.9.0 resolved the advisory accepted at v0.8.0 (dependency status).
Technical report: Bastion: Execution Boundaries, Control Evolution, and Bounded
Evaluation of a Governed Agent Fleet (10 November 2026, 25 pages), DOI
10.5281/zenodo.24120596. The PDF, source manifest
and source snapshot are attached to the report-2026-11 release; their hashes are
in the publication manifest. Sources are in report/.
The v0.9.2 source code is archived on Zenodo as software, DOI
10.5281/zenodo.24121343.
It is the one definitive report; the 6–7 November PDFs were drafts, not editions. The fleet was retired on 9 November 2026 (D4): its serving
resources were deleted, the Agent Gateway is retained because a retained Runtime engine
references it, and billing was disconnected from the project. The evidence above is dated
history; see HANDOFF for the verified end state.
The public console redeploys from
main, by maintainer decision: merging a PR publishes its site/ changes.
ADK session state does not cross A2A in either direction. Bastion therefore carries findings and routing decisions as explicit message content and requires confirmation from the step that did the work (ADR-012).
Checked 6 November: issue #6854 is closed as completed. Community PR #6859, opened by Sylvester Kaczmarek and including Arjun Ganesh's legacy-path warning and end-to-end test, was imported as d2cb22e on 12 October. The GitHub PR was closed after import, not merged through GitHub. Both warnings first shipped in v2.12.0; the state boundary remains.
Use CPython 3.12; on Windows the launcher must know that interpreter, otherwise invoke its full path. No GCP credentials are needed for the offline suite.
py -3.12 -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements-dev.txt -c requirements.lock
ruff check .
ruff format --check .
mypy agents gateway identity registry runtime model_armor observability infrastructure
pytest tests --cov --cov-report=term-missing --cov-fail-under=100
python scripts/check_docs.py
python scripts/check_versions.py
python scripts/render_diagrams.py --check
python scripts/redact_evidence.py --check
node site/build.mjs
python scripts/check_site.py
npx markdownlint-cli2 "**/*.md"The Python core suite has 398 tests and 100% configured statement/branch coverage. Offline synthetic experiments and report tooling are outside that production core scope.
Live work requires approved credentials, explicit project/quota values and the credits-only conditions in AGENTS.md. The operator's default CLI project is not Bastion's. The access prerequisites describe the required roles.
.\infrastructure\bootstrap.ps1 `
-Project 'YOUR_PROJECT_ID' `
-MemoryAgentEngineId 'YOUR_MEMORY_ENGINE_ID' `
-RuntimeAgentEngineId 'YOUR_RUNTIME_ENGINE_ID'Bootstrap provisions and deploys infrastructure and runs a live smoke test; it is not an
offline verification command. A code-only Orchestrator update requires both Cloud Run
(infrastructure/deploy.sh) and python -m infrastructure.deploy_agent_runtime.
Cloud Run/data transport use europe-north2; managed agent services, screening and audit use
europe-west4; Gemini uses Vertex AI global. Model processing has no end-to-end EU residency
claim. Field handling and retention are in Data governance.
Firestore owns admission, leases, attempts and idempotency. Completed replays do not repeat work; failed attempts remain retryable. Pub/Sub's five-attempt dead-letter target is approximate: E21 records one rejected event's five deliveries, not a guaranteed maximum or worker-timeout sequence. Cloud Run max instances is a scaling limit, not a monetary cap.
Operations documents objectives and dry-run-first rollback/teardown. Teardown preserves Firestore, secrets, Runtime/Memory and audit state by default. Residual resources and costs need an owner decision before applying the agreed end state.
GitHub deployment uses Workload Identity Federation, no long-lived GCP key, and a manually dispatched workflow. Its configured authorization boundary is distinct from an observed deployment through that CI path; no such successful CI deployment is claimed here.
- Architecture, Security, Data governance
- Report plan, claim matrix, audit, remediation ledger
- Observation/defect history, evidence, site instructions
- Architecture decisions — index: 001 real IAM · 002 three agents · 003 managed platform · 004 global Gemini · 005 ADK and A2A · 006 observable pillar proof · 007 tool poisoning · 008 human approval loop · 009 Model Armor threshold · 010 policy enforcement gate · 011 inbound screening covers tool results · 012 structured findings across A2A
Bastion can request review but cannot change IAM. It makes no legal-certification, immutable audit-storage, historical SLO or wall-clock-week-test claim. The audit bucket is retained but unlocked; configured alerts are not proof of delivered paging.
The hackathon archive, gallery and fixed video retain their submission dates. No award is claimed. Do not re-render embedded architecture assets during routine cleanup.
Personal research software, without a support or patch SLA. Code is under the MIT License; the technical report and the evidence records are under CC BY 4.0.