What it does
WEXSPACE AI — Human-Governed Agentic Core handles a narrow professional workflow: reconcile synthetic receiving evidence, explain the discrepancy, and prepare a review package for a human. Deterministic tools perform the calculation. Durable state records completed work. The agent stops at HUMAN_REVIEW and has no approval or rejection tool.
How it works
- Compile the professional request and check required inputs.
- Use the Strands Agents SDK to expose inspection, reconciliation, and review-package tools.
- Persist each tool result and execution cursor in SQLite.
- Record a hash-linked evidence ledger and deduplicate recorded effects.
- Export the decision package and evidence for human review.
- After interruption, reload the same work ID and continue from its saved cursor.
Repeated named Strands tool calls return their own saved results. They do not advance a different workflow step. Missing inputs remain visible to the agent.
Latest public verification
The public repository is the reproducible implementation. GitHub Actions installs it into a fresh Ubuntu environment from the public repository, checks dependencies, executes the full test suite, runs the README workflow, performs a real SIGKILL/fresh-process resume test, and records hashes.
Current public main commit: 46de154b4c1ca684a77b78f1218435757670395a
Latest successful public replay: https://github.com/moatazfarea/wexspace-human-governed-agentic-core/actions/runs/34857936134
That run verifies:
- fresh public-source installation with no broken Python requirements;
- Strands Agents SDK 1.55.1 installed and direct named-tool execution exercised;
- HUMAN_REVIEW enforcement;
- real SIGKILL followed by a fresh Python process;
- cursor recovery from 1 to 3;
- one recorded effect for each completed tool;
- no duplicate side effects;
- evidence-chain integrity;
- README synthetic workflow reproduction.
The interruption test kills after the durable cursor is published. It establishes recovery for this local workflow; it does not establish exactly-once behavior for an external distributed transaction.
Strands and AWS status
The source contains a native Strands Amazon Bedrock route using the standard AWS credential chain. Live Bedrock model inference and AgentCore deployment/invocation remain separate qualification gates and are not claimed as completed by the public deterministic CI run.
An environment-configured AgentCore bridge is present in source. A deployed AgentCore runtime, independent-provider model failover, and a public live backend are not claimed as completed capabilities in this entry unless separately verified before final submission.
Additional qualified fleet evidence
The public repository also contains isolated qualification code for LangGraph durable recovery and official Model Context Protocol reference servers. These capabilities are kept distinct from the core Strands path and do not replace Strands as the hackathon foundation.
Architecture
Request compiler → Strands tool interface → deterministic professional tools → SQLite cursor and effect ledger → hash-linked evidence → human review package.
Source, setup instructions, tests, architecture notes, and MIT license: https://github.com/moatazfarea/wexspace-human-governed-agentic-core
Safety and prior work
WEXSPACE is an ongoing company-building program. This entry uses its general governed-execution ideas; the public Git history records this implementation and subsequent changes. Examples use synthetic data. No employer-confidential records, credentials, private provider identifiers, or private WEXSPACE governance files are included.
The tools cannot finalize a consequential decision. Human approval remains outside the agent tool surface.
Log in or sign up for Devpost to join the conversation.