AegisData · Theater
see how the firewall thinks, one tool call at a time
🛡️ What's an ATV? (click to expand the 30-second concept)
Every action an AI agent attempts (read a file, run a shell command, query a DB) is captured as a fixed 2,080-dimensional vector — the Agent Trace Vector. The firewall has 5 cheap checks (300-340) that look at different parts of that vector and decide ALLOW / REQUIRE_APPROVAL / BLOCK. Below, you'll watch a simulated agent take one step at a time. When something goes wrong you'll see which check caught it and what would have happened without the firewall.
scenario
1 / 13
🤖
Agent says
"…"
about to call
📄
tool
{}
Aegis runs 5 checks
Verdict
🛡️ …
If allowed, would have…
📄
…
ATV-2080-v1 (which band Aegis examined)
📍 Source-code paths — who calls Aegis, and where the verdict is computed
YOUR AGENT HOST
gathers inputs from
pre-LLM → post-LLM → pre-tool
ask_aegis() / hook ⇒ POSTs
POST
→
/evaluate
AEGIS
build_atv → 5 checks → sign
① ② ③
JSON
→
verdict
YOUR AGENT HOST
acts on verdict
run / block / approve
Note: the host does not build the 2080-D vector itself — build_atv() in Aegis does. The host's job is to assemble the inputs (tool_name, args, plan, state, safety_flags, cost_estimate, memory_fp) across multiple lifecycle stages, then POST them in one shot. See the "Host-side instrumentation map" card below for which ingredient is captured where.