Agent Management Guide
Step-by-step guide to configuring, running, and monitoring ACE's 10 autonomous agents
Before agents can run their observe/reason/act/reflect loop, they need access to an LLM. ACE supports Anthropic (Claude), OpenAI (GPT-4), and Google (Gemini).
# Option 1: Set via Dashboard
Dashboard → Settings → AI Configuration → Add API Key
# Option 2: Set via environment variable
export ANTHROPIC_API_KEY="sk-ant-..." # or OPENAI_API_KEY, GOOGLE_API_KEYDashboard
Open the Agent Command Center to see all 10 agents with live status, last run time, and error counts.
MCP Tool
ace_agent_status()# REST API — get all agent statuses
curl http://localhost:7777/api/v1/agents/status \
-H "Authorization: Bearer $ACE_JWT_TOKEN"
# Response shows each agent's enabled state, last run, and error count# Trigger the Knowledge Guardian
curl -X POST http://localhost:7777/api/v1/agents/guardian/run \
-H "Authorization: Bearer $ACE_JWT_TOKEN"
# Trigger via MCP
ace_trigger_agent({ agent_name: "guardian" })# Update Knowledge Guardian config
curl -X PUT http://localhost:7777/api/v1/{namespace}/agents/guardian/config \
-H "Authorization: Bearer $ACE_JWT_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"enabled": true,
"interval_hours": 2,
"llm_provider": "anthropic"
}'
# Via MCP
ace_agent_config({
agent_name: "guardian",
enabled: true,
interval_hours: 2,
llm_provider: "anthropic"
})Configurable Options
enabled— turn agent on/offinterval_hours— how often it runsllm_provider— anthropic, openai, or google
Cost Management
Lower intervals = more LLM calls = higher cost. Start with default intervals and adjust based on the cost tracking in the Command Center.
The Knowledge Guardian, Risk Engine, and Pattern Detector surface actionable findings. Review these regularly in the Insights Panel.
# Get agent insights
curl http://localhost:7777/api/v1/{namespace}/agents/insights \
-H "Authorization: Bearer $ACE_JWT_TOKEN"
# Via MCP
ace_agent_insights()Guardian
Contradictions, stale data, knowledge drift
Risk Engine
Actionable risks, causal chains, temporal patterns
Pattern Detector
Recurring patterns in work logs and decisions
# View last 10 runs for the Guardian
curl "http://localhost:7777/api/v1/{namespace}/agents/guardian/history?limit=10" \
-H "Authorization: Bearer $ACE_JWT_TOKEN"
# Each run includes a full trace:
# - observe: what entities were scanned
# - reason: what the LLM decided (with token count)
# - act: what actions were taken
# - reflect: confidence score and self-evaluationAfter a bulk data import
Trigger graph_maintenance → then consolidator → then guardian. This validates relationships, deduplicates memories, and checks for contradictions.
Start-of-day routine
Check ace_agent_insights() for overnight findings. Review any contradictions flagged by the Guardian. Clear resolved items.
Before a code review
Trigger quality_governor to check for incomplete entities, then risk_engine to surface any risks related to the changes.
Reduce agent costs
Increase intervals for Discovery agents (6h → 12h), keep Protection agents frequent. Use ace_agent_health() to monitor token spend per agent.