Why We Bet on Autonomy Over Assistants

There are two ways to build AI products.

Assistants: AI helps humans do tasks. Human stays in control.

Autonomous agents: AI does tasks. Human defines goals.

We bet on autonomy. Here's why.

The Efficiency Gap#

Every human-in-the-loop is latency.

Assistant Mode:
1. AI: "I found 47 relevant documents. Which should I analyze?"
2. Human: [waits 3 hours to respond]
3. AI: "Here's my analysis. Want me to draft something?"
4. Human: [waits 2 more hours]
5. AI: "Draft complete. Should I send it?"
6. Human: [finally available]
Total time: 8 hours

Autonomous Mode:
1. Human: "Analyze relevant docs and send summary to team"
2. AI: [does all of that]
3. Human: [receives notification when done]
Total time: 12 minutes

Same outcome. Different orders of magnitude.

When Autonomy Makes Sense#

High-Volume, Low-Stakes#

Processing 10,000 support tickets for routing.

Nobody wants to approve each one. Approve the process, not the instances.

Clear Success Criteria#

"Book me a flight to NYC under $400 on Tuesday."

Either it's booked correctly or it's not. Human judgment at each step adds nothing.

Speed-Sensitive#

Customer waiting for a response. Server needs scaling. Market window closing.

Humans are the bottleneck. Remove them.

Expertise Bottleneck#

One expert. Thousands of decisions that need their judgment.

Train the AI on the expert's patterns. Scale the wisdom.

When Autonomy Doesn't Make Sense#

High-Stakes, Low-Reversibility#

Firing someone. Major financial decisions. Public statements.

Humans should own these. AI provides input, not decisions.

Novel Situations#

First time seeing a problem. Edge case with no precedent.

AI excels at pattern matching. Novelty breaks patterns.

Relationship-Dependent#

"Client wants a discount."

This isn't analytical. It's political, historical, relational. Keep humans here.

Stay Updated

Get updates on new labs and experiments.

The RAG Reality Check: Why Retrieval Isn't Magic

RAG is everywhere, but production implementations fail constantly. Common failure modes, debugging strategies, and what actually works in retrieval-augmented generation.

The Trust Ladder#

Autonomy isn't binary. It's a spectrum.

Level	AI Does	Human Does
0	Nothing	Everything
1	Suggestions	Decisions + Actions
2	Drafts	Approval + Actions
3	Actions	Approval only
4	Actions + Notifications	Exception handling
5	Everything	Goal setting

Most AI products are Level 1-2. We build Level 3-5.

Building Trust Through Transparency#

Autonomy without transparency is dangerous.

Every autonomous action should leave:

What happened: The action taken
Why: The reasoning that led to it
With what confidence: Uncertainty quantified
How to undo it: Reversibility path

Users don't need to approve everything. They need to understand everything.

The Human Role Shifts#

When AI does tasks, humans do:

Goal setting: What should happen? Exception handling: What to do when things go wrong? Judgment calls: Ambiguous situations that need wisdom. Relationship work: Things that require human connection.

Less doing. More directing. More human work becomes more human.

Our Approach#

Every Kingly agent is designed for maximum useful autonomy:

Clear scope: What it can and can't do
Transparent reasoning: Why it did what it did
Graceful escalation: When to involve humans
Audit trails: What happened, always

The goal isn't replacing humans. It's freeing humans for human work.

The Future We're Building#

In 5 years, asking AI a question and waiting for it to ask you questions back will feel... quaint.

"Why is this AI bothering me? Just do it."

The products that win will be the ones that figure out useful autonomy first. Not as a feature. As a philosophy.

That's the bet.

2026 Field Notes: Closing the Action Gap#

Traditional APIs are no longer enough. The industry is rapidly shifting towards Vision-LLM scaffolding to close the "Action Gap." In our work with NAAC building the COPI (Co-Pilot Intelligence) module for the experimental Tarragon aircraft, we've replaced rigid state machines with a hybrid neural-symbolic architecture.

Furthermore, we're leveraging SOFIA (Self-Organizing Flight Intelligence Agent)—our open-source RL framework with 42 training levels, 15 RL algorithms, 29 aircraft configs, and 128 parallel environments—to train autonomous agents through domain randomization rather than hard-coded heuristics.