ACB Delegara

Delegate boldly. Approve with proof. clearly communicates autonomy plus governance.

Comment

Inspiration

Every startup founder, CTO, and engineering lead knows the cognitive overload of release day: reviewing dozens of open PRs, checking architecture specifications, verifying test suites, inspecting security scans, and compiling evidence for stakeholders.

When teams turn to generative AI agents for delegation, they are immediately confronted with two dangerous extremes:

  1. Unbounded, Reckless Autonomy: Agents that hallucinate permissions, bypass critical gates, or auto-approve destructive production deployments without human oversight.
  2. Fragile Toy Scripts: Shallow chatbots that lose context, crash on tool failures, or convert missing evidence into artificial success.

We built ACB Delegara on an uncompromising principle: Bounded autonomy with truthful failure propagation. Delegara acts as a tireless executive delegation agent—handling the heavy lifting of coordinating specialists, running deterministic checks, and gathering cryptographic evidence across developer tools—while strictly halting for human review before any sensitive action.

What it does

ACB Delegara transforms high-level founder objectives into verifiable, audit-backed execution:

  • Targeted Intake & Clarification: If a founder’s objective is ambiguous (e.g., omitting target versions or service scopes), Delegara halts immediately and asks clarifying questions instead of guessing.
  • Typed DAG Planning: Synthesizes declarative, strongly-typed assignment DAGs governed by versioned recipes.
  • Specialist Sub-Agent Execution: Dispatches specialized workers across distinct domains:
    • Architecture Specialist: Audits architectural notes and checks mandatory security boundary specifications.
    • QA Specialist: Executes test suites, validates semantic versioning, and performs secret scanning.
    • Evidence Specialist: Reconciles collected receipts against formal acceptance criteria.
    • Delivery Specialist: Automates issue tracking across Jira Cloud, Linear, and Slack Block Kit.
  • Deterministic Policy Enforcement: Evaluates pure-logic safety action classes (READ_SAFE, WRITE_SCOPED, SENSITIVE) with zero LLM hallucination risk. Sensitive actions (publishing release tags, modifying databases) are deterministically blocked until explicit human approval.
  • Cryptographic Receipt Stream: Every tool call produces an append-only, boundary-redacted receipt (rcpt_*) stored in Google Cloud Firestore.
  • Standalone Review Packets (Invariant I-12): Compiles all evidence, diffs, and plain-English findings into a standalone Review Packet with founder decision controls (Approve & Apply Changes or Reject).
  • Interactive Multi-Channel Decision Gates: Founders can review and authorize actions directly via the Web Command Center, Slack Block Kit action buttons, or mobile magic links.

How we built it

ACB Delegara is built on a Dual-SDK Hybrid Google Agent Architecture:

  1. Root Orchestration Agent: Powered by the Google Antigravity SDK using Gemini 3.1 Pro via Vertex AI for high-reasoning DAG composition, founder intent evaluation, and objective clarification.
  2. Specialist Sub-Agents: Powered by the Google Agent Development Kit (ADK) using Gemini 3.7 Flash for fast, cost-effective subagent execution across Architecture, QA, and Delivery tasks.
  3. Deterministic Policy Gate: A pure Python policy layer (never consulting an LLM) enforcing non-negotiable boundaries before side effects occur.
  4. Full-Stack Application: FastAPI 0.115 backend paired with Jinja2, HTMX, and Tailwind CSS for a sub-millisecond, reactive Executive Command Center.
  5. Infrastructure & Persistence: Google Cloud Run multi-stage deployment (non-root unprivileged container), Google Cloud Firestore Native for multi-tenant append-only storage, and automated Terraform provisioning.
  6. Integrations: Direct connectors for GitHub, Slack, Jira Cloud, Linear, Notion, and Google Drive.

Challenges we ran into

  • Preventing Agent Sycophancy: LLM agents instinctively want to report success even when underlying tools fail or evidence is missing. We solved this by designing a Golden Demo service (fixtures/demo-service) with four deliberate seeded defects (missing security section, failing scanner adapter, ambiguous input, prohibited sensitive mutation). We engineered Delegara's evidence reconciler to strictly propagate failure states into the Review Packet rather than masking them.
  • Forward-Compatibility Tenancy: Enforcing strict multi-tenant isolation across all records. We established 12 non-negotiable architectural invariants (I-01 to I-12) and backed them with dedicated regression tests.
  • Sub-Millisecond Policy Guarantees: Ensuring policy evaluation adds zero noticeable latency to agent workflows by enforcing pure-logic set operations and static action classification without network or model round-trips.

Accomplishments that we're proud of

  • 100% Quality Gate Pass Rate: 444 automated unit, integration, and invariant tests passing with 0 lint errors (ruff) and 0 type errors across 204 files (mypy strict).
  • Extreme Cost & Latency Efficiency: Average workflow execution cost of just $0.081 USD with sub-second execution overhead.
  • Truthful Failure Reporting: Delegara is proven to catch seeded defects and recommend REJECT when criteria are unmet, giving founders genuine confidence.
  • Production Infrastructure: Fully containerized on Google Cloud Run with automated Terraform provisioning and zero secret leakage verified by deterministic scanners.

What we learned

  • Truthful failure is vastly more valuable than artificial success: In mission-critical software and startup leadership, a delegation agent that accurately flags missing evidence saves hours of post-release firefighting.
  • Separating planning from policy is essential: Generative models excel at planning and synthesis, but safety, permissions, and side-effect gating must remain deterministic.

What's next for ACB Delegara

  • Multi-Agent Consensus Verification: Enabling multi-agent cross-examination where specialized agents critique each other's evidence before human presentation.
  • GitHub Action & CI/CD Native Runner: Triggering Delegara directly inside GitHub Actions workflows for automated PR sign-offs.
  • Enterprise Multi-Workspace RBAC: Expanding tenant organization hierarchy with delegated team roles (Requester, Reviewer, Admin, Auditor).

Built With

Share this project:

Updates

Submission history