Your AI agents.
Under protection.
WatchMyAgents is a runtime security infrastructure for autonomous AI agents. Three layers — Watch, Guardian AI and Shield — connected by a live feedback loop that turns observability into adaptive enforcement, agent by agent.
Ready to protect your IT FORTRESS ?
AI agents are the new attack surface.
LLM + tools + actions introduce risks classic security stacks were never designed to see.
Data leaks
Prompt injection, exfiltration via tools, verbose logs leaking secrets and PII.
Bad operational practices
Plaintext secrets, excessive permissions, no traceability across runs.
Behavioral drift
Loops, action escalation, abnormal cost spikes, out-of-scope actions.
Compromise & corruption
Hijacked tools, malicious instructions, agents taken hostage.
Recursive Fractal Security Loop™
Watch. Guardian. Shield. A self-reinforcing loop where every observation feeds smarter analysis, every analysis feeds stronger policies, and every policy sharpens the next observation — on each agent, then on whole teams of agents.
Watch
Instruments and collects every agent execution trace — building a reliable timeline of what each agent tried and what it actually did.
- Model calls, tool calls, data access, sensitive actions
- Real-time alerts: Info → Warning → High → Critical
- Triage context for every signal, per agent & per env
Guardian AI
The intelligence layer. Normalizes Watch signals, scores risk and hygiene, governs the lifecycle of rules and proposes ready-to-validate policies for Shield.
- Signal correlation, hygiene & risk scoring per agent
- Auto-suggested rules with rationale & false-positive estimate
- Governance: simulation, approval, versioning, rollback
Shield
The enforcement engine — per-agent, per-environment policies that protect each agent from itself and from the outside world.
- Tool allow/deny lists, parameter & domain restrictions
- Injection / exfiltration protection, PII & secret redaction
- Rate limits, token budgets, loop caps, auto-quarantine
The same loop, at every scale
The Watch → Guardian → Shield loop runs recursively — on a single agent, on a team, on a full multi-agent system, and across your whole organization.
Per-agent observation, scoring and policy enforcement.
Cross-agent correlation inside a legion (e.g. Support, Finance).
System-wide risk patterns across agents, tools and data flows.
Org-level posture, shared policies and collective intelligence.
One loop, three layers
Watch observes. Guardian AI thinks. Shield enforces. Each layer is built to do one thing, and to do it brilliantly.
The all-seeing eye of your fleet.
Watch instruments every agent at the SDK and tool layer. Model calls, tool calls, parameters and data access are captured, classified and turned into a triage-ready signal — without ever leaving your machine.
Full-spectrum capture
Model calls, tool calls, parameters and sensitive data access — instrumented at the SDK layer, nothing slips through.
Real-time severity triage
Every signal classified Info → Warning → High → Critical, with the context needed to act in seconds.
Tamper-evident timeline
An immutable, audit-ready record of what every agent tried and what it actually did, per agent and per environment.
Local-first runtime
Runs on your machine alongside your agents. Zero telemetry leaves your perimeter without your explicit consent.
The reasoning core of the loop.
Guardian is the brain between Watch and Shield. It interprets signals, reasons about intent, and turns raw telemetry into precise, explainable policy decisions — at machine speed, under human authority.
Contextual reasoning
Correlates Watch signals across agents, tools and time windows to surface real threats — not noise.
Adaptive policy synthesis
Drafts new Shield policies on the fly from observed behavior, ranked by impact and false-positive risk.
Human-in-the-loop
Every suggestion is explainable, simulatable and reversible. You stay in control of what ships to prod.
Continuous learning
Feeds policy efficacy back into the loop — Guardian gets sharper with every incident across the fleet.
The enforcement perimeter of your agents.
Shield runs alongside your agents and applies Guardian-approved policies in real time. Tool allowlists, parameter restrictions, rate limits, PII redaction, auto-quarantine — every guardrail is enforced before damage happens.
Tool allow / deny lists
Per-agent, per-environment restrictions on tools, parameters and domains. Block dangerous calls before they fire.
Injection & exfiltration block
Stops prompt injection, secret leakage and PII exfiltration in real time, with automatic redaction on the wire.
Rate & token budgets
Hard rate limits, token caps and loop detection — runaway agents are throttled or quarantined automatically.
Approved by Guardian
Every active rule is versioned, simulated and signed off. Roll back any policy in one click without redeploying agents.
How it works
Watch and Shield run locally on your machine. Guardian runs in the cloud on anonymized data. You stay in control.
Observes your agents and records a daily log.
Nothing is deployed without your explicit approval.
Applies the approved policies on your machine.
YOUR FORTRESS
Your AI agents. Under protection.
One command center to watch every action, score every risk, and enforce every policy — with full audit trail.
Sentinel.Knight · on watch
Guardian inbox
Protected agents
- Assistant Personnel CEOANTHROPIC-MANAGED · active · last seen 31/05 22:15:18
- Agent FinancierANTHROPIC-MANAGED · active · last seen 30/05 15:04:34
- Deep researcherANTHROPIC-MANAGED · active · last seen 30/05 13:04:46
- Test AgentANTHROPIC-MANAGED · active · last seen 25/05 20:25:04
- agent_01UNy3MizTnJ3s7Wg…ANTHROPIC-MANAGED · active · last seen never seen
Quick actions
- ShieldManage policies
- GuardianReview suggestions
- WatchTail signals
- KeysManage API keys
Validation queue
RULE_ID agent-financier-deny-new-tools
{
"tool_name": { "not_in": ["bash", "write"] },
"action_type": "tool_use"
}Live timeline
- 15:03:40bash· (default)
- 15:02:45bash· (default)
- 15:02:40write· (default)
- 16:45:26web_fetch· URL outside curated allowlist
- 16:42:32web_search· (default)
- 16:42:31web_search· (default)
Shield · Policies
5 active · 4 from Guardian- deep-researcher-websearch-high-error-denyDeny web_search if error rate is highDEEP RESEARCHERDENY
- deep-researcher-webfetch-error-alertAlert on Web Fetch errorsDEEP RESEARCHERINTERRUPT
- p2-webfetch-allowlistweb_fetch allowlistFLEETDENY
- agent-financier-new-tool-bash-denyDeny new tool · bashAGENT FINANCIERDENY
Reports & Audit
Every decision auditable. Exportable for SOC2 evidence.
Threat Intel
Live IOCs, adversary playbooks, agent-specific feeds.
Compliance
SOC2 · ISO27001 · EU AI Act mapping out of the box.
How to install it
Watch it all in 3 steps. Literally 30 seconds.
Packages available for
Wrap, run, observe
Initialize once with your agentId. Every prompt, tool call and response is now streamed to your Fortress in real time.
- 1Framework API keyThe key your application / framework uses to connect to its main provider (Anthropic, OpenAI, …).
- 2Watch My Agent API keyThe key tied to your Watch My Agent account.
- 3Agent ID from your framework consoleThe unique identifier of the agent you created in the framework console.
Watch sees everything. Guardian AI thinks. Shield stops the rest.
Every agent class. Every risk profile.
Coding agents
Tool-using agents writing, deploying and altering code.
Customer support
Agents touching PII, tickets and customer history.
Internal ops
Finance, HR, ops agents with access to sensitive systems.
Multi-agent platforms
Swarms, orchestrators and complex agent workflows.
Command your LEGIONS.
Organize agents into squads by team — Customer Services, HR, Marketing, Dev Team. Apply policies, monitor hygiene and orchestrate the whole fleet from a single command center.
Privacy by design.
Your logs stay inside your information system. Only encrypted, anonymized or pseudonymized signals are forwarded to WatchMyAgents — never raw PII, secrets or business content. On serious threats, deeper investigation is requested through the channels you define.
Don't wait for your first agent incident.
Join the early-access program and deploy Watch + Shield on your production agents in days, not quarters.
Or email minedor@watchmyagents.com