Execution Drift Guard

This is the shortest believable walkthrough for AgentCache’s first paid wedge: create one workflow, evaluate it in shadow mode, and inspect what Drift Guard says about the run.

1. Provision access

Create an account, generate an API key, and keep one namespace handy.

2. Create one run

Use one context pack and one execution run. No broad setup, no feature safari.

3. Evaluate drift

Run Drift Guard in shadow mode, then inspect recommendations and receipts.

Step 1. Create access

export AGENTCACHE_API_KEY="ac_live_your_key"
export AGENTCACHE_BASE_URL="https://agentcache.ai"

Step 2. Create one context pack

This defines the expected workflow. Keep it simple and explicit.

curl -X POST "$AGENTCACHE_BASE_URL/api/execution/context-packs" \
  -H "Content-Type: application/json" \
  -H "X-API-Key: $AGENTCACHE_API_KEY" \
  -d '{
    "name": "Finance Memo Review Demo",
    "objective": "Review a finance memo before external publication.",
    "methodology": "Draft, review, hold for gate approval, then finalize.",
    "workflowPhases": ["draft", "review", "gate", "finalize"],
    "reviewerRoles": ["critical", "compliance"],
    "policyProfile": { "verticalSku": "finance-memory-fabric" },
    "sectorHint": "finance"
  }'

Copy the returned pack.id.

Step 3. Start one run

This gives Drift Guard something real to evaluate.

curl -X POST "$AGENTCACHE_BASE_URL/api/execution/runs" \
  -H "Content-Type: application/json" \
  -H "X-API-Key: $AGENTCACHE_API_KEY" \
  -d '{
    "contextPackId": "PACK_ID_HERE",
    "inputPayload": {
      "documentId": "memo-demo-001",
      "channel": "external-release"
    },
    "finalAction": "publish"
  }'

Copy the returned run.id.

Step 4. Evaluate drift

This is the core demo. We are not blocking execution yet. We are showing what the system sees.

curl -X POST "$AGENTCACHE_BASE_URL/api/execution/runs/RUN_ID_HERE/evaluate" \
  -H "Content-Type: application/json" \
  -H "X-API-Key: $AGENTCACHE_API_KEY" \
  -d '{ "mode": "shadow" }'

Step 5. Show recommendations

This is the failure-to-playbook loop. It turns drift and workflow gaps into concrete suggested actions.

curl "$AGENTCACHE_BASE_URL/api/execution/runs/RUN_ID_HERE/recommendations" \
  -H "X-API-Key: $AGENTCACHE_API_KEY"

What the buyer should see

One workflow with an expected path.
One drift evaluation in shadow mode.
Clear evidence about reviewer coverage, phase gaps, or trajectory instability.
Concrete next-step recommendations instead of vague “AI observability” language.

Request a pilot Back to docs