AEGIS · Agentic Infrastructure · Base • 2025–26
When agents act autonomously, someone still has to answer for it.
AEGIS is the supervision layer for autonomous DeFi agents. As sole designer through Base Batches 003, I designed the interface that gives protocol teams real-time visibility into what their agents are doing, why they're doing it, and the controls to intervene when they shouldn't be.
AEGIS · Planning Preview · Helios awaiting approval
4:47 until executionAgents
3 / 3
Capital
$2.4M
Interventions
12
Uptime
99.8%
The Problem
Web3 teams are deploying AI agents that rebalance positions, execute liquidations, and manage treasury around the clock. The moment a team deploys one, nobody knows what it's doing or why. They find out when TVL drops or a panicked Telegram message lands at 3am. Execution in DeFAI is solved. Oversight isn't.
Execution Timeline
The execution timeline is a live feed of what's executing, what's waiting for review, and what's been overridden. Colored left edge communicates status before a word is read.
Execution Timeline · All agents
Live↑ Click any row to expand reasoning. Every decision reconstructable.
Planning Preview
Before an agent executes, it surfaces the full plan: move size, price impact, risk factors, alternatives considered. There's a countdown. The clock makes the stakes real.
Planning Preview · Helios
Awaiting approvalRebalance $2.4M USDC — Morpho to Compound
✓ Slippage tolerance: 0.15%
✓ Gas ceiling: 25 gwei
↑ Interactive — approve, modify, or reject. The countdown is live. Inaction is a decision.
Guardrails
Guardrails define when an agent can act without asking. The principle: readable by a risk manager, not just an engineer.
Guardrail Status · All agents
1 breach↑ Green: nominal. Amber: approaching. Red: breached. No color is decorative.
Override & Alerts
When an agent breaches a guardrail or behaves anomalously, someone needs to act fast. The override layer gives operators three levels of intervention: pause a single agent, pause all agents, or emergency stop. The design challenge was making these controls accessible without making them easy to hit accidentally. A misclick on emergency stop is its own kind of incident.
Override Controls
1 pausedAnomaly & Alerts
3 unackedExplainability
The explainability layer reconstructs the full decision chain: why this action, what alternatives were considered, what the confidence level was. That's what trust actually looks like: earned confidence, not blind faith.
Explainability · Nemesis · Decision reconstruction
94% confidenceNemesis detected volatility and executed protective collateral increase.
Design Decisions
Countdown over hard deadline
Three iterations to get this right. No countdown: agents executed unreviewed. Hard deadline: operators panic-rejected. Escalating countdown: operators started checking earlier, calmly. The timer shapes behavior more than any button on the page.
Human language over parameter syntax
Guardrails were originally config values. A risk manager who couldn't parse them was the turning point.
Explainability added late
Without reasoning traces, AEGIS was a control panel. With them, operators stopped reacting to outcomes and started understanding the system.
Outcomes
Three protocol teams in testing, each with a live oversight failure.
Anomaly detection time in live testing, down from "whenever someone noticed."
Testing teams named Planning Preview the screen they couldn't work without.
Protocol teams in live testing with real agent failures and real capital.
Agent interventions triggered during testing. 4 automated, 8 operator-initiated.
“For the first time I feel like I actually understand what our agents are doing. Not just what they did, but what they're thinking.”
— Protocol operator, Base Batches 003
Reflection
The first version of the Planning Preview had no countdown. Operators treated it like a notification. They'd check it eventually. Agents executed before anyone reviewed. The second version had a hard deadline with a red timer. Operators panicked and rejected plans they should have approved. The third version, the one that shipped, uses a calm countdown with escalating visual weight. Operators started checking earlier without the stress response. That calibration took three iterations and taught me something about designing for autonomous systems: the interface isn't just showing what the agent will do. It's shaping how the human responds. Get that wrong and the oversight layer becomes either ignored or adversarial.