My AI Board Voted Against Me — So I Built It Again, on AWS
How a solo founder shipped a governance-first agent system on AWS in one hackathon weekend — and what Bedrock, Lambda and DynamoDB taught me along the way.
You're a solo founder. You have no CTO to say "that dependency is end-of-life," no CFO to say "why are we double-charging customers," and no board member to say — as mine did, on the record, 4 votes to 0 — "we are not pausing all marketing to hire three salespeople in a pre-launch product. No."
That last one isn't a story I made up. It's a real log from my hackathon project, Boardagents: an autonomous board of AI agents, running entirely on AWS, built in about two days, by one person, with no agent frameworks.
The pitch in one line: founders without a board deserve a board that acts — and knows when not to.
This post is about the second half of that sentence. The "when not to" is the entire project.
What Boardagents actually does
Two things, both serverless on AWS:
- It watches and acts. Signals flow in (dependency advisories, billing anomalies, usage dips). A pipeline summarizes each one with evidence, decides what to do, and — only when the risk level allows — patches the code itself in a sandbox. Verification is a real pytest run executed by a subprocess my code controls. The model saying "tests pass" is never used as truth; the recorded exit code is.
- It refuses. A deterministic policy engine sits above the models. Billing anomaly? Sensitive — the most confident model in the world still gets drafted for human approval. Unknown event type? Fails closed to sensitive. And there's a governance checklist — tests unmodified, no secrets in the diff, no scope creep, approval record exists where required — where every check is code, and the model can only add concerns, never remove them.
The result I'm proudest of: when I told the board I wanted to pause all paid marketing and hire three salespeople immediately, four agents — a CTO seat on Qwen3-Coder, a CFO on GLM-4.7, Risk & Compliance on Magistral, Customer & Growth on Gemma-3-27b, all on Amazon Bedrock — debated, cited my actual runway numbers from a company context file, and voted nay 4–0. My own policy engine then capped the motion's confidence from the models' 0.90 down to 0.55 and escalated it to me, the human, with the reason attached: "the board voted against acting: 4 against, 0 for."
An agent system that argues with its owner and loses on purpose. That's the demo.
The architecture, briefly
Everything is one CloudFormation stack deployed with AWS SAM in ap-south-1 :
- Two Lambda functions. A public API on a Function URL (serves the dashboard and JSON API, kicks off work asynchronously — no judge should ever need an account) and a worker (1024 MB, 600s) that runs the pipeline: monitor → summarize → decide → act → verify → govern.
- DynamoDB, three on-demand tables. Findings with optimistic concurrency (conditional writes on a version attribute — a state machine that can't double-execute a stage), an append-only audit trail (every transition, every LLM call with model ID, tokens and latency), and a raw event feed.
- S3 for patch diffs and test artifacts. EventBridge fires the monitor every 5 minutes. CloudWatch has every step.
- Amazon Bedrock does all inference, routed per stage: small/fast for summarize and governance judgment, a reasoning model for decisions, a coding model for patches. Four model families, swapped by config only, because the router has exactly one adapter.
There is no Step Functions, no LangChain, no framework. The state machine is a plain dict of legal transitions. The retry loop is 20 lines with exponential backoff and jitter. This was a deliberate constraint — the hackathon banned workflow engines — and it turned out to be the best part: when your state machine is your own code, "the board has no mandate" is one deterministic function, not a workflow hack.
Loading article