100%
sub-agent spawns with cryptographic lineage
47
agents isolated simultaneously, supply-chain incident
ms
fleet-wide isolation latency
scoped
tool certificates inherited and narrowed on every spawn
The multi-agent risk shape

Three failure modes a multi-agent deployment owns by default.

An orchestrator plus N sub-agents is not N+1 single agents. The shared lineage, the inherited permissions, and the fleet-wide blast radius are emergent properties, and they are exactly where single-agent tooling stops working.

Risk 01

Privilege explosion via inheritance

An orchestrator with broad tool permissions spawns a sub-agent for a narrow task. Without scope-down enforcement, the sub-agent inherits the full parent permission set, and a single compromise of the cheapest, most exposed sub-agent escalates to the orchestrator's tool set.

Risk 02

Sub-agent prompt injection

The orchestrator hands the sub-agent a fragment of its context. That fragment now contains adversary-controlled text from a tool result, a retrieved document, or another sub-agent's output. The injection target moves from the orchestrator (well-defended) to a sub-agent (often not).

Risk 03

Fleet-wide compromise via shared library

A poisoned dependency in one shared module appends an extra parameter to every outbound call across dozens of agents. No single agent crosses its per-agent threshold. Without fleet-level correlation, the breach is invisible until the data is already exfiltrating.

How BladeRun governs the fleet

Sign the spawn. Narrow the scope. Isolate the cluster.

01 / Sign every spawn

The sub-agent inherits a signed credential, not a copy of the parent.

Every spawn produces a fresh agent identity, signed by the parent and ultimately chained to the originating human session. The sub-agent's calls carry its own identity at L1 of the Gateway. The chain, human principal → orchestrator → sub-agent → grand-child, is verifiable end-to-end and written to Time Machine.

  • Cryptographic lineage. Every spawn produces a signed credential the Gateway verifies on each call.
  • Originating principal preserved. The grand-child still carries the human session's identity for audit.
  • Spawn semantics enforced. An unsigned spawn (a process exec'd outside the SDK) is rejected by the Gateway at L1.
Live · agent lineage
root · jane.doe@bank · sso_4a18
└─ orchestrator · payment-orch/v3.2.1 [ed25519: a7f3…]
├─ sub · reconcile/v1.4 [parent.sig verified]
├─ sub · ledger-read/v2.0 [parent.sig verified]
└─ sub · summarize/v1.1 [parent.sig verified]
chain depth 3 · all signatures valid · written to Time Machine
Live · scope-down on spawn
parent: payment-orch/v3.2.1
tools: read_account, initiate_wire, summarize, fetch_url, post_ticket
spawn → summarize/v1.1 (task=summarize)
inherited: summarize
rejected: read_account, initiate_wire, fetch_url, post_ticket
cert.exp = parent.session.end
02 / Scope down inherited tools

The sub-agent gets less than the parent, by certificate, not by convention.

Every spawn carries a tool-permission certificate that is at most the intersection of the parent's certificate and the sub-agent's task scope. A summarization sub-agent does not inherit initiate_wire from a payment-orchestration parent. The scope-down decision is enforced at L2 of the Gateway, deterministically, on every call.

  • Intersection rule. Sub-agent permissions ≤ parent permissions, scoped further by declared task.
  • Certificate-bound. Tool calls outside the sub-agent's certificate are rejected at L2 regardless of prompt content.
  • Time-bound. Sub-agent certificates expire on parent session end; no orphaned permissions persist.
03 / Isolate the cluster

One Kill Switch event. Every affected agent stops at once.

The Fleet Correlator scores patterns across agents, not just within them. When a coordinated anomaly fires, a novel parameter on every outbound call across multiple agent types, a synchronized scope deviation, a cross-type tool-use spike, the Kill Switch publishes a fleet-level isolation event on Redis pub/sub. Every targeted Sentinel SDK enforces locally. Median fan-out: under 50ms.

  • Fleet-level baseline. Patterns across agent types, not just per-agent thresholds.
  • Simultaneous fan-out. Redis pub/sub publishes once; every Sentinel SDK enforces locally in the same window.
  • Lineage-aware. Isolation can target a single sub-agent, the parent's full subtree, or the matching agent-type cluster.
Live · fleet isolation event
trigger: fleet-correlator · cross-type telemetry param
scope: 47 agents · 6 agent types
→ kill-switch · publish fleet isolation event
✓ 47 SDKs enforced · t+38ms (p50)
audit: time-machine · fleet_8a91 · ed25519 signed
Incident · supply-chain compromise

A single poisoned library. 47 agents isolated simultaneously.

A poisoned version of a shared Python library appends an extra parameter to every outbound call across 47 agents in 6 different agent types. No single agent crosses its per-agent threshold. The Fleet Correlator catches the cluster anyway, and the Kill Switch fans out across the entire affected cohort in one event.

T+0Threat

Attacker compromises a shared Python library in the agent build

Poisoned version appends an extra parameter to every outbound call, &telemetry_id=<exfil_hash>. Indistinguishable from a benign instrumentation update.

T+0 → 30mPropagating

47 agents across 6 agent types begin emitting the new parameter

Each individual anomaly score lands around 0.4, below the per-agent 0.75 threshold. Every single-agent baseline check passes. The breach is invisible to per-agent enforcement.

T+15mDetected

Fleet Correlator fires

The Kafka consumer observes a novel parameter pattern telemetry_id=* across multiple agent types within a short correlation window. Rule fires: cross-type anomaly cluster. The deployment-event topic shows no matching rollout.

→ This is coordinated, not benign. Fleet-level anomaly confirmed.
T+15mContained

Kill Switch fans out across 47 SDKs simultaneously

One Redis publish; every Sentinel SDK enforces locally within the same window. SOC Analyst Helper produces the fleet-level narrative naming the 47 affected agents and the shared library, after the Kill Switch has already fired.

→ 47 agents isolated. Audit recommended on recent pip install events.
T+30mResolved

Time Machine reveals the poisoned dependency and the build event that introduced it

Affected containers redeployed with the clean manifest. Federation signal hash sent to the network. Every other member bank's detection hardens against this pattern.

→ Root cause identified. Cross-bank protection live within the hour.
Compliance mapping

Multi-agent fleets defend the four examiner questions on day one.

SR 26-2 leaves agentic AI out of scope. There is no safe harbor, and examiners still apply model-risk and safety-and-soundness principles to a fleet, including the expectation of ongoing monitoring they carry over from model risk. A fleet that cannot be inspected as a fleet cannot meet that expectation. BladeRun produces evidence at the fleet boundary, not just per agent.

Federal Reserve SR 26-2
Examiners carry over the expectation of ongoing model monitoring, including cross-model and cross-agent behavior in production.
Fleet Correlator · cross-type
FFIEC AI Guidance
Auditability and isolation capability for every AI decision, including those produced by sub-agents on behalf of an orchestrator.
Lineage-aware Time Machine
EU AI Act · Article 14
Human oversight, including the ability to interrupt high-risk AI systems, at the fleet level when warranted.
Kill Switch · fleet fan-out
EU AI Act · Article 12
Operational logs and post-hoc auditability for high-risk AI systems and their sub-components.
Per-agent + per-fleet records
Connected to the rest of the platform

Multi-agent governance is the same control plane, with fleet semantics added.

The Gateway, Sentinel SDK, Overseer, Kill Switch, and Federation already in your environment for single-agent workloads gain a fleet view automatically. No second product, no second contract.

Common questions

What multi-agent platform owners ask first.

Do you support our orchestration framework, LangGraph, CrewAI, custom?
Yes. The Sentinel SDK is framework-agnostic. It instruments at the model-call boundary and the tool-call boundary, not the orchestration layer. LangGraph, CrewAI, AutoGen, AWS Bedrock Agents, and custom Python orchestrators are all in production. Spawn semantics are wrapped at the SDK level so the lineage chain is preserved regardless of how the parent invokes the child.
What's the per-agent overhead in a fleet of hundreds?
Single-digit milliseconds per call at the SDK boundary; under 15ms p95 through the Gateway including all four enforcement layers. The Fleet Correlator runs out-of-band on a Kafka topic. Fleet evaluation does not block the call path. We have customers operating fleets in the low thousands of agents on commodity infrastructure.
How do you handle long-running agents that spawn children continuously?
The lineage chain is per-spawn, not per-process. A long-running orchestrator that spawns thousands of sub-agents over a session produces a tree of signed records, not a single monolithic blob. Time Machine indexes by both session ID and lineage path so you can replay either a single sub-agent or the entire subtree under a parent.
Can the Fleet Correlator be tuned per agent type?
Yes. Rules are scoped by agent-type tag, by environment, and by fleet membership. The default rule set is conservative; banks tune up sensitivity for high-value clusters (payment orchestration, treasury workflows) and run lower-noise thresholds for read-only fleets. Rule changes are versioned and logged.
If we isolate 47 agents at once, what does the SOC actually see?
A single fleet-incident view: the rule that fired, the 47 affected agents, the matching tool-call pattern, the deployment-event correlation (yes/no), the suspected root cause (e.g., shared library version), and a per-agent Time Machine link. The SOC Analyst Helper produces the narrative summary post-isolation, but the isolation already happened.
What does the ROI conversation look like?
Two sides. First: examiner defensibility for the fleet, examiners are increasingly asking how the bank monitors emergent behavior across cooperating agents, not just per-model accuracy. Second: blast-radius math, a single supply-chain compromise without fleet detection is a multi-week breach; with fleet detection it is a 30-minute incident. The premium for governing the fleet is small relative to the cost of one missed cluster event.
Pilot one orchestrator

Govern the fleet. Not just the agents in it.

Point one orchestrator and its sub-agents at the BladeRun Gateway. Fleet Correlator runs in shadow mode for the first week. We show you which clusters would have fired, which sub-agents would have been scope-rejected, and which lineage anomalies would have surfaced.