HEADLINE FEATURE · REPLAY A POLICY CHANGE & SEE THE DIFF
TRACE: run-aegis-commerce
⬇ DOWNLOAD TRACE (.JSONL)
QUICK POLICY PRESETS:
1. Strict ($20 Gate + Depth=2)
2. Lax (Allow Delete + $500 Gate)
3. Unfenced Writes (Lost Writes 0→3+)
4. Default Policy
Press "▶ REPLAY & DIFF" to re-run the recorded trace against the edited policy and compare outcomes.
Seq
Agent
Tool & Payload
Before Decision
After Decision
Rule
Click "▶ REPLAY & DIFF" above to see every call whose decision changed.
CHAOS MODE · EXPOSED vs. GATED UNDER FAULTS (0% · 10% · 30%)
⚡ RUN CHAOS SWEEP
Inspired by Netflix Tech Blog ("The Netflix Simian Army" / Chaos Monkey). Injects latency spikes, HTTP 500s, malformed JSON, 429 storms & partial writes.
Loading chaos survival chart...
STEP 1 & 4 · RECORDED TRACE (.JSONL) & EXPORT TO REGRESSION TEST
✓ RUN EXPORTED TESTS
Seq
Agent
Request
Gateway Decision
Upstream
Action
Loading recorded JSONL trace...
HIDDEN-AGENT AUTO-FINGERPRINTER (UNMASKS BACKEND AGENTS WITHOUT X-AGENT-ID HEADERS)
⚡ SIMULATE HIDDEN SWARM (ZERO-KEY)
↻ REFRESH
CLEAR
1. Wrap Any Local Agent App (Downloadable CLI)
python -m gateway.cli inspect --mock -- python examples/hidden_swarm.py
2. Or Point 1 Env Var at Local HiveGate Proxy (Zero Code)
OPENAI_BASE_URL=http://127.0.0.1:8010/v1/proxy/gemini
Fingerprint ID
Inferred Hidden Role
System Prompt Preview
Tools (Declared / Emitted)
Calls & Burst
Risk Flags
No hidden agents captured on local proxy yet. Click "⚡ SIMULATE HIDDEN SWARM (ZERO-KEY)" or run hivegate inspect.
Wave
Agent
Endpoint
Exposed (Pink Pass)
Agent Gateway (Teal Pass)
Run an audit to record the 6 intercepted multi-agent waves.
Run an audit to inspect the 6-agent hierarchy and Groq planning rationales.
GRADE: NOT EVALUATED (RUN AUDIT)
DETECTED STACK: not evaluated
LATENCY WATERFALL: not measured (run audit to measure)
Pillar & Evaluation Dimension
Measured Current State
What It Was Lacking (Before)
What It Has To Be (Target Benchmark)
Status & Module
Load any URL or press RUN AUDIT to compute the 4-Pillar Evaluation & Benchmark Scorecard.
1. Proxy Env
COPY
AGENT_GATEWAY_URL=http://127.0.0.1:8010
UPSTREAM_URL=http://127.0.0.1:8010/targets/aegis-store
2. YAML Policy
COPY
budgets:
max_calls: 120
max_depth: 3
read_only_agents:
- reader-audit-01
3. Python SDK
COPY
from sdk.client import GatewayClient
gw = GatewayClient("http://127.0.0.1:8010")
await gw.call(goal_id="g1", agent_id="a1", tool="read")
README · PROBLEM, HOW IT WORKS & MEASURED PROOF
Problem: When an agent swarm hits a backend concurrently, duplicate reads trigger 429 Too Many Requests, naive concurrent writes overwrite each other (lost updates), recursive sub-agents loop without bound, and read-only agents invoke destructive endpoints.
How HiveGate works: Singleflight coalesces duplicate in-flight reads, leased locks + monotonic fencing tokens serialize concurrent writes, call-tree budgets prune recursion at max_depth=3, and declarative YAML rules block or pause risky actions.
Measured proof (Aegis Storefront): Exposed pass: 52 HTTP 429 overload errors, 3 lost writes (final inventory=1 instead of 8), 2 destructive actions ran unchecked. Gated pass: 0 429s, 0 lost writes (final=8), 44 backend calls avoided, recursion halted at depth 3, and 95.8% survival under 30% Chaos Monkey fault injection.
HOW AGENT GATEWAY PROTECTS YOUR BACKEND (5 ACTIVE GUARDRAILS)
IN-PROCESS CONTROL PLANE
01 / ADAPTIVE ADMISSION & FAIR QUEUE (admission.py · scheduler.py)
CAP: 3 CONCURRENT · 0 429s
Caps concurrent upstream requests and schedules goals via weighted fair queuing. When 36+ agents burst at once, excess calls wait in a bounded queue instead of triggering 429 Too Many Requests.
02 / SINGLEFLIGHT DEDUPLICATION & TTL CACHE (dedup.py)
COALESCED: -- · SAVED: --
Hashes (tool, normalized_payload) so identical in-flight read requests share a single upstream call, and caches recent reads to eliminate redundant LLM sub-agent polling.
03 / FENCED LEASES & VERSIONED CAS STATE (locks.py · state.py)
LOST WRITES: 0 · FENCED
Grants expiring leases with monotonically increasing fencing_token numbers and requires If-Match version checks on shared state so concurrent writers never overwrite each other.
04 / GOAL BUDGETS & RECURSION DEPTH BREAKER (budgets.py)
MAX DEPTH: d=3 · PRUNED: --
Tracks parent→child call_id lineage across the entire agent swarm and halts runaway recursive fan-out (`budget_exceeded: max_depth`) before exponential token/API cost explosions.
05 / YAML POLICY ENGINE & HUMAN APPROVAL GATE (policy.py)
DELETE BLOCKED · >=$100 GATED
Enforces per-agent tool rules (blocking read_only_agent from DELETE endpoints) and pauses high-value financial actions (`>= $100`) in pending_approval until human sign-off.
TEST INDIVIDUAL GUARDRAIL:
1. Thundering Herd
2. Singleflight Dedup
3. Fenced Locks
4. Recursion Cap
5. Policy & Approval
MEASURED PROTECTION TELEMETRY
EXPOSED vs. GATEWAY
Protection Metric
Exposed
With Gateway
Valid Call Completion
--
--
Client Calls Issued
--
--
Backend Calls Sent
--
--
Calls Avoided (Saved)
0
--
Singleflight Deduplicated
0
--
Upstream 429 Overload Errors
--
0
Peak Backend Concurrency
--
--
Lost Concurrent Writes
--
0
Recursion Depth Reached
--
--
Runaway Calls Pruned
0
--
Policy Violations Blocked
0 (Allowed!)
--
Latency (p50 / p99)
--
--