Core Agent¶

The Core Agent is the unified default entry point for OLAV — ask a question, get an answer. It handles 80% of daily network queries in one shot, and automatically delegates to specialised sub-agents (writer, ops) when the task requires it.

v0.15.0 — Core Agent replaces Quick Agent

As of v0.15.0 the former Quick Agent workspace has been merged into Core Agent. All capabilities are preserved. The olav command now routes through core by default.

Feature Claims

ID	Claim	Status
C-NE-15	`olav "R1 BGP neighbors?"` returns query results from snapshot data	✅ v0.10.0
C-NE-16	Blacklisted commands are rejected by `execute_cli`	✅ v0.10.0
C-NE-17	`/learn_cmd "show version" --device R1` executes and generates a TextFSM template	✅ v0.13.1
C-NE-18	`take_snapshot` triggers incremental collection with a new snapshot_id	✅ v0.13.1
C-NE-19	Core Agent suggests upgrading to NetOps Agent for complex multi-step analysis	✅ v0.13.1
C-V15-01	`olav "<question>"` routes to core agent by default (no `--agent` flag required)	✅ v0.15.0
C-V15-02	Semantic router selects specialised agent when confidence > 0.85	✅ v0.15.0

Legend: ⚠️ (E2E-gated) = test exists but requires env-var flag, skipped in CI; ⚠️ (manual) = demo-verified only, no automated CI test

When to Use Core Agent¶

Task	Core Agent	NetOps Agent
"How many BGP peers are down?"	✅ Yes	Overkill
"Show R2 interfaces"	✅ Yes	Overkill
"What changed since yesterday?"	✅ Possible	Better (use Diff)
"Simulate link failure impact"	❌ No	✅ Yes (Analysis)
"Deploy a test lab"	❌ No	✅ Yes (Lab)

Core Agent handles simple queries by generating SQL against the snapshot database or executing a single CLI command. For anything requiring iteration, comparison, or simulation, it will suggest using the NetOps Agent — or route there automatically if semantic confidence is high.

Query Examples¶

Natural Language to SQL¶

olav "Are all BGP neighbors established on R1?"

OLAV translates this to:

SELECT neighbor_ip, neighbor_as, state
FROM v_bgp_neighbors_auto
WHERE device_name = 'R1' AND state != 'Established'

Direct SQL¶

olav "execute_sql: SELECT device_name, count(*) as cmd_count FROM netops.parsed_outputs GROUP BY 1"

CLI Execution¶

olav "Run 'show ip route' on R2"

The agent looks up the command in the whitelist, connects via SSH, and returns the output. Blacklisted commands are rejected before any connection is made.

Snapshot Collection¶

Trigger an on-demand snapshot:

olav "Take a snapshot of R1 and R2"

This runs the collection pipeline for the specified devices, producing a new snapshot_id. The data is immediately queryable.

Learning New Commands¶

When a show command has no TextFSM template, use /learn_cmd to teach OLAV:

olav "/learn_cmd 'show ip bgp community' --device R1"

This will:

Execute the command on the device via SSH
Analyse the output structure
Generate a TextFSM template
Save it to workspace/ops/config/templates/custom/

The template is available for all future collections without restart (hot-reload).

Tools Available¶

Core Agent has access to these tools (injected from netops/tools/ via inject_into_core):

Tool	Purpose
`execute_cli`	Run a single CLI command on a device
`search_commands`	Find available commands by platform and keyword
`diff_configs`	Compare device configs between two snapshots
`take_snapshot`	Trigger on-demand collection

Plus all core platform tools (execute_sql, format_and_export, olav_recall_memory, etc.) which are always available.

Escalation¶

When Core Agent determines a question requires multi-step analysis, it tells you:

"This question requires comparing multiple data sources across time. I recommend using the NetOps Agent: olav --agent netops 'your question'"

This happens for questions involving simulation, drift detection, or multi-device correlation. With v0.15.0 semantic routing enabled, this handoff can happen automatically.