Autonomous AI SRE that learns your system and fixes it. Drop a manifest. Claude builds the integrations. Neo does the rest.
# describe your environment
$ cat neo.yaml
vcs: github | issues: jira | alerts: pagerduty
metrics: prometheus | deploy: jenkins | infra: aws
# claude builds the integrations
$ claude "bootstrap neo from neo.yaml"
Generated: github provider, jira provider, pagerduty alerter...
# neo observes, investigates, fixes, learns
$ ./agent-neo
[neo] sweep #1: 847 signals → 3 novel → 0 incidents (1.2s)
[neo] t2: sub-agent service-health: service running, transient restart
[neo] learned: suppress optimizer_service_down (from #42)
Three Tiers, One Agent
Progressive escalation. Small model handles triage. Medium model investigates with focused sub-agents. Large model reviews code and creates fixes. All serialized, all auditable.
Triage
Fast model observes signals, applies learned patterns, files issues. Most signals handled without LLM.
Investigate
Sub-agents with 1-3 tools each. Focused questions, structured answers. Go makes the decision.
Fix
Full agentic code review and fix. Creates PRs with investigation context. Optional auto-merge.
Self-Learning
Neo starts with zero knowledge. Every resolved issue teaches it a new pattern. Patterns are JSON files tracked in git — deterministic, auditable, no retraining needed.
Pattern Engine
Signals matched against learned patterns before any LLM call. Same signal twice? Handled in microseconds, not minutes.
Signal Diffing
Only novel or changed signals reach the LLM. 1,400 signals per sweep → 4 novel. Sub-second sweeps.
Evidence Validation
If the LLM says "service down" but the tool output says "active" — contradiction caught, incident dropped. No hallucinated issues.
Training Data
Every investigation stored in a vector database. Past similar signals retrieved before LLM. Export for fine-tuning.
Quick Start
1. Create neo.yaml
Describe your VCS, issues, alerts, metrics, deploy, and LLM providers.
2. Bootstrap
$ claude "bootstrap neo from neo.yaml"
3. Deploy in NOOP mode
Neo observes and files issues. Doesn't touch anything. You close issues and Neo learns.
4. Disable NOOP
Neo creates PRs. You merge. Auto-deploy via CI. Neo verifies. Regression? Rollback to previous tag.
5. Full autonomy
Enable auto_merge. Neo reviews its own PRs, merges approved ones, deploys, and monitors. You sleep.
Pluggable Everything
Neo doesn't care if you use GitHub or GitLab, PagerDuty or Slack, AWS or bare metal. Describe it in neo.yaml, Claude builds the integration.