agentvcs holds an agent’s whole history — every iteration, with a quality score attached. That lets it answer things a live runtime can’t: is the loop improving or quietly rotting? Is a branch worth keeping? Is a shared memory safe?
Two narrated, runnable walkthroughs. Both use real evaluations under the hood — nothing here is faked.
Five everyday situations anyone running an AI product recognizes — a support bot that gets a little worse every week, a client fork nobody merged back, paying for context that changes nothing, one bad fact poisoning a fleet. Each ends in a one-line recommendation, with no math on screen.
Read the five stories →The same five capabilities, shown with the real numbers and the theory behind them:
the Price equation catching an error catastrophe, Muller’s ratchet on an isolated
branch, the information value of context in bits, and branching-process containment.
Every claim is asserted against --json.
| Situation | Command | The call it makes |
|---|---|---|
| A self-updating agent quietly getting worse | agentvcs price | Stop — the loop is net-negative; roll back |
| The same agent, branch-and-keep-the-best | agentvcs price | Keep going — it’s genuinely improving |
| A fork maintained alone for too long | agentvcs branch | Merge it back before it rots further |
| A prompt stuffed with context that changes nothing | agentvcs infobits | Trim the context — cut cost & latency |
| A shared memory that could spread a mistake | agentvcs contain | Verify N% of reads, or shrink the fleet |
Run them yourself:
bash examples/business-cases/run.sh ·
bash examples/evolution-diagnostics/run.sh —
full guide in docs/DEMOS.md.