Multi-agent orchestration
The six-agent swarm
Six specialists share one blackboard with a battle log: parallel dispatch, phase-aware skill hints and cross-phase negotiation. Supervised, target-locked, and fully audited. For authorized targets only.
The specialists
Six agents, one shared mission
Each agent owns a phase of the assessment. They read and write shared state instead of passing messages in a chain. For authorized targets only.
recon
Recon
Attack-surface discovery and fingerprinting
Scans, fingerprints, and scores the attack surface. Opens the run and keeps the target model fresh as new services appear.
vuln
Vuln
Vulnerability / CVE and module correlation
CVE and exploit correlation plus module matching, pairing threat intel (NVD, EPSS, KEV) with applicable capabilities.
exploit
Exploit
Capability selection, payload work and execution
Selects modules, crafts and mutates payloads, and executes through the MCP tool layer under policy and allowlist.
post_exploit
Post Exploit
Credential, loot and lateral-target handling
Handles credentials and loot, and generates lateral targets inside the same target lock.
- Safety gate
critic
Critic
Scope, risk and policy review
Reviews scope, risk and policy before execution. Forbidden actions and disallowed assets die here, before anything runs.
- Learns
reflection
Reflection
Strategy review and lessons
Reviews strategy and records lessons learned in the experience store so the next run starts smarter.
How a run flows
The swarm lifecycle
One loop around the swarm: discover, correlate, pass the critic gate, execute, consolidate, learn, then loop with a smarter plan. Every transition is written to the battle log.
Phase 1 · Discover
Recon
Fingerprint the authorized target and score the attack surface. Seeds the blackboard target model.
Phase 2 · Correlate
Vuln
Match threat intel and CVEs to applicable capabilities. Raises hypotheses, not payloads.
Gate · Review
Criticmust pass
Reviews scope, risk and policy. Forbidden actions and out-of-scope assets are killed here. Nothing executes without a pass.
Phase 3 · Execute
Exploit
Selects modules, crafts payloads and executes through the policy-enforced tool layer.
Phase 4 · Consolidate
Post Exploit
Handles credentials and loot, and proposes lateral targets inside the same target lock.
Phase 5 · Learn
Reflection
Reviews strategy and records lessons to the experience store, so the next loop starts smarter.
The loop repeats until the plan is exhausted or the operator stops it. Reflection output feeds the next recon pass, so coverage builds up instead of restarting from zero.
Why it stays safe
Critic and reflection: the oversight pair
Two agents keep the other four honest. They carry no payloads, only veto power and memory.
Critic reviews before execution
- Checks every proposed action against mission scope and the allowlisted target lock.
- Kills forbidden actions and disallowed assets before anything runs.
- Flags high-risk steps for operator confirmation instead of silently proceeding.
Reflection learns after execution
- Reviews what worked, what failed, and what the plan missed.
- Writes lessons to the persistent experience store for future runs.
- Shifts strategy mid-run, for example steering recon toward newly exposed services.
Blackboard: shared truth
The target model, hypotheses, findings and plan live in one shared state store. Agents read and write the blackboard instead of chaining messages, so parallel recon and vuln research never diverge.
Battle log: full audit
Who decided what, and when. The battle log feeds the tamper-evident audit chain, so every run is debuggable, resumable and reviewable. See the safety model.
Run modes
Swarm mode or campaign mode
One swarm, two ways to drive it. Same target-IP lock and permission model either way.
Single-target swarm mode
--swarm splits one authorized target across the six specialists: parallel recon plus vuln research, critic pre-check, reflection shifts. It fits going deep on a high-value system you own or have written permission to test.
Persistent campaign orchestration
Without --swarm, the autonomous campaign drives a persistent multi-phase queue (recon, exploit, privesc, lateral, validation) with adaptive aggression, resume and checkpoints across one or many authorized targets.