Newshot

Dual-System Decision-Making + Two-Level Self-Evolution: Agnes Speeds Up Agents by Up to 14x and Cuts Costs to 1/27

AI summary

Agnes Harness has been open-sourced.

Translation(model translation, not the source's own text)

Dual-System Decision-Making + Dual-Layer Self-Evolution: Agnes Makes Agents Up to 14x Faster and Slashes Costs to 1/27!

Agnes Harness has been open-sourced

By Yishui, from QbitAI | WeChat Official Account QbitAI

In the past, building a set of enterprise software might cost 90 units of effort to develop, and only 10 units to adapt it for each customer. Build the software once, sell it to many customers, and that initial 90 gets amortized bit by bit. That's the business logic of traditional software companies. AI has changed this formula. Now development only costs 1, but adaptation still costs 10.

If you know, you know — this calculation that Zaniel, an independent FDE (Forward Deployed Engineer) and Stanford PhD student, shared with QbitAI earlier almost perfectly captures the most painful reality of Agent deployment today.

Development costs plunged from 90 to 1, but that stubborn 10 — the hardest nut to crack — is still sitting there, completely unmoved.

Because enterprise adaptation, at its core, is about getting AI to re-learn every customer's business processes, tool systems, and on-site equipment from scratch.

Yet many Agents today are essentially "cyber temp workers" trapped inside browsers and terminals.

They write code with dazzling flair, and their Benchmark scores are prettier than each other.

But the moment you ask them to step outside the screen — to command a conveyor belt on the factory floor, read the output of an industrial weighing scale, or operate a robotic arm to complete a precision assembly — the picture changes immediately.

Sorry, I haven't seen this before. You'll need to teach me again.

The classic trope of "you only get work experience once you're on the job" — I get it, so let's teach it carefully.

But after all that effort to teach it, it turns around and gives you a "fish memory" performance —

In the same customer scenario, the stupid mistakes it makes on the 100th task execution are almost identical to the ones it made on the 1st.

Tokens burn round after round, but not a single bit of experience is retained.

So how do you get an Agent to quickly adapt to new tasks while also accumulating execution experience?

Agnes Harness.

An open-source system licensed under Apache 2.0, permitted for commercial use, and already boasting 2.1k Stars on GitHub — that's how it caught my attention.

Give the Agent fast and slow brains to drive down the cost of getting work done

Let me highlight the key points first — the two most noteworthy things about Agnes Harness are:

One is the Sys1 dual-system decision-making, and the other is TTT × RSI dual-layer self-evolution.

By understanding these two innovative designs, you'll see exactly where the bottleneck in Agent deployment lies today.

To make the whole process easier to follow, let's just dive right into a game of "Rebirth" (doge).

You wake up one day and suppose you're now the owner of a factory. To cut costs and boost efficiency, you have just one thing on your mind —

I want to bring AI into my factory.

Looking at the market, the models are getting stronger one after another, but when it's actually time to pay up and deploy, the first hurdle appears:

How on earth do I drive down the execution cost of each task?

No matter how sexy the AI concept is, it can't end up costing me money, right? I need to carefully calculate the ROI first.

Being the smart you, you decide to check the books first — where was all the money going before?

After a thorough audit — good grief — the biggest expense turns out to be LLM API calls.

The reason many Agents today are slow and expensive isn't that the LLM isn't powerful enough. On the contrary, it's because it's too capable — so everything gets dumped on it.

Many operations that don't need complex reasoning at all still require a full LLM call. Isn't that just pure overkill and waste??

Agnes Harness thought — no, we need division of labor — so it went ahead and built a "Sys1 dual-system decision-making" mechanism.

For the parts that need complex reasoning, hand them to the large language model System 2.

For the high-frequency, well-defined operations, hand them to a specially trained 9B decision model System 1 — that is, Agnes Sys1.0.

It can be plugged into Agnes Harness as a plugin and is专门 designed to judge what to do next:

Which tool to call? Keep waiting or execute immediately? When unsure, should the LLM be brought in?

Don't underestimate this 9B model — its "solo combat capability" is quite formidable.

According to the Agnes Harness report, in a simulated home water leak task, when judging what to do next, the LLM took an average of 7.6 seconds, while Agnes Sys1.0 needed only 0.53 seconds.

That's directly 14.3x faster than the original.

As for cost, in another set of tests in a luggage room (completing item weighing, vertical placement, and temporary storage), each side made 100 judgments. The online LLM's per-call cost was approximately $0.00247, while Agnes Sys1.0's was approximately $0.000091.

By that math, the per-decision cost is only about 1/27 of the former.

With this "ultimate assistant," once the two systems work together, wouldn't it just take off??

Agnes Harness actually brought in a robot dog and tested it in a simulation environment.

The robot dog's task was to find a door indoors and go through it.

Compared to calling the LLM at every step, the "LLM + Agnes Sys1.0" duo directly reduced overall task time by 32.9%.

Just looking at the numbers might not give you a feel for it, but click on the video and compare — good grief, the speed difference is truly visible to the naked eye.

Speed goes up, and cost comes down.

But what if you switch equipment today and connect a new task tomorrow — Agnes Sys1.0 hasn't seen it before?

You can't exactly drag the model back for retraining every time a new job comes in.

Fortunately, Agnes Sys1.0 has another trick up its sleeve — TTT, which stands for Test-Time Training.

Stop having amnesia every day — let task experience actually stick

How do you get an Agent to quickly adapt to new tasks? A natural idea is:

But here's the problem — traditional model fine-tuning requires first collecting data for the new task, then training and deploying.

The whole pipeline takes considerable time and cost.

TTT is more interesting — it directly has the Agent do a crash course before going on duty.

Agnes Sys1.0 first looks at what the new task involves, then generates a few practice problems with standard answers for itself, and after completing them, adjusts its judgment approach based on the results.

The entire process doesn't touch the main parameters of the 9B model — it only fine-tunes a small set of calibration parameters at the very end.

However, this "crash training" has a rather interesting design choice:

Huh? You just barely learned it — why forget it again??

After all, today you're operating equipment A, tomorrow it might be equipment B. If everything learned gets stuffed into the model each time, who knows whether they'll start fighting each other over time.

So Agnes Sys1.0 simply makes temporary adjustments to a small subset of parameters for the task at hand each time, and restores everything to original after finishing.

Sounds reasonable, but as a factory owner, I can't help worrying:

Model parameters can be restored, but the hard-earned experience from all that exploration — surely that can't just go down the drain too, right?

That's where the other half of "dual-layer self-evolution" comes in — RSI.

RSI, Recursive Self-Improvement — everyone's familiar with this by now.

After each job is done, first hold a retrospective meeting for the Agent.

Agnes Harness reviews the entire execution process, looking at where errors kept recurring and which steps wasted the most time, then adjusts the coordination of Prompt, Memory, Skills, and Subagent.

If a better method is discovered, it's first validated within the current task.

Once confirmed effective, it's then organized into a versioned Playbook for direct reuse in subsequent tasks.

Agnes Harness actually tried this out on a complex equipment inbound inspection task.

Simply put, when equipment arrives at the warehouse, the Agent needs to follow the procedure to operate the weighing scale, read the weight, and complete the inspection.

The first time, the Agent crashed and burned, as expected...

It rushed to read the scale before it stabilized, didn't figure out the operation sequence properly, fumbled around for a full 30 minutes, and ultimately timed out.

But RSI recorded the pitfalls from this attempt.

As a result, the second time around, it directly followed the validated procedure.

Note that the new methods summarized by AI must first go through simulation testing and be confirmed reliable before they can be written into the Playbook.

So, you're allowed to improve yourself, but don't experiment on my equipment. (doge)

At this point, the Agent is finally no longer a "cyber temp worker" starting from zero every day.

But no matter how proficient it gets inside the screen, when it arrives at the factory floor, it can't exactly tighten screws with a mouse, right?

Getting the Agent out of the screen — that's what's needed now.

Step out of the screen — give the Agent a body that connects to the real world

However, giving the Agent a body isn't about slapping on a robotic arm right away.

To actually work in a factory, it first needs to interface properly with existing business systems.

Orders, warehousing, approvals... every factory has its own set of processes and tools.

You can't possibly rebuild and hack a whole new Agent for every new customer, right??

Agnes Harness's approach is quite pragmatic:

Business tools, model adapters, Agent workflows, Skills, and even the operation interface — all can be encapsulated as plugins.

Existing business systems can also be connected directly through MCP.

That sounds a bit abstract, so I just tried it out myself.

Agnes Harness doesn't have a ready-made online web version yet — to try it, you need to deploy it locally first.

The barrier to entry isn't too high — just prepare Node.js 24.10 or above and pnpm 10.34.5. If you're using cloud model APIs, no additional GPU configuration is needed.

Just throw the GitHub address directly at DeepSeek Harness and let the AI install AI for me. (doge)

I went on typing on the side, and about 50 minutes later, Agnes Harness was running on my MacBook Air.

You can use the demo model provided by the project, or configure your own model API.

I directly connected the model I usually use.

Next, it was time for something practical.

I casually put together a proofreading plugin to check for typos, grammatical errors, inconsistencies, and number and unit issues.

Not long after I finished describing the requirements, the plugin appeared in the pending review list.

After manually confirming that the code and permissions were fine, I clicked enable, and the status immediately changed to "Local" and "Ready."

I casually threw in a draft with deliberate traps.

Good heavens, it actually caught 6 issues, even marking the location, issue type, and suggested fixes.

Even more interesting, when I opened a new conversation window and simply said "help me proofread," the Agent automatically invoked this plugin.

No need to re-explain the checking rules every time.

The plugin also has a version management mechanism: new sessions can use the new version, while tasks already in progress continue using the original version.

Now the benefits of pluginization are quite intuitive.

What was installed today is a proofreading tool; in a factory, it could be order queries or warehouse approvals.

You thought that was it? Big no, no, no — the real hard nut to crack is still hardware.

Robotic arms, robots, sensors — how are they supposed to understand the Agent's arrangements?

Agnes Harness's approach is to define a unified set of "communication rules" for these devices, called MHS (Model Hardware Standard).

After a device connects, it first "introduces itself" in a unified format:

what state I'm currently in, which actions I can perform, and what data I can return.

As for who to call first, who to call next, and how to break down the task, that is left to Agnes Harness to coordinate.

The official team also made a Mars rescue demo.

After receiving the rescue mission, the Agent first dispatches search devices to find the astronaut, then coordinates the robot and the airlock to bring the person safely back to the base.

Find the person, open the door, move, return.

Multiple devices relay work around the same goal, and this time the Agent is finally not just clicking a mouse on a screen.

But when physical devices are really involved, safety is no joke.

Agnes Harness is responsible for arranging tasks, while real-time control and physical safety remain the responsibility of the device controller.

Currently, MHS is being integrated into the new version of Agnes Harness as an official plugin.

At this point, the Agnes Harness system has been walked through from head to toe.

The large language model is responsible for thinking, Agnes Sys1.0 is responsible for quick judgments, TTT adapts to the new task at hand, RSI retains long-term experience, and plugins and MHS connect software systems with real devices.

Good heavens, with this combination punch, how do the results actually look?

The official team released an interim self-assessment scorecard covering 9 benchmarks.

Across 9 evaluations, 6 took first place against the control group, and another 1 tied for first with Codex.

Specifically, DeepPlanning, which tests complex task planning, gave Agnes Harness 87.07 points; BigCodeBench-Hard V1, which tests complex coding ability, reached a pass rate of 40.69%.

With these results laid out, we factory owners finally have something to look forward to.

In the future, when paying AI to do the work, at least we won't have to clean up the mess behind it every day. (doge)

Make every delivery the starting point for the next delivery

At the beginning of the year, bosses were still in FOMO, afraid their companies would be one step slow in adopting AI.

Now, looking at discussions online again, the tone has changed somewhat.

Some people have started complaining that when an AI Agent completes the first 80% of a task, it feels like magic, but the last 20% can leave them debugging for a full 4 hours.

Others have it worse. A SaaS practitioner said he charges clients $400 a month, but a buggy Agent frantically retried 31,000 times in less than 6 hours, burning $4,700 in tokens overnight and leaving him more than ten times in the red.

Even executives in the AI industry can't take it anymore.

Cyera Chief Strategy Officer Sol Rashidi previously revealed that she had already laid off half of her Agents.

The time spent watching AI work was more than the time AI saved her.

That's a bit awkward...

All I can say is that after the tide goes out, what everyone ultimately looks at is still the economics.

From an Agent's entry to operation, maintenance, and then replication, every step can generate a new bill.

What Agnes Harness is trying is exactly to push these four bills down, one by one.

Saving money on a single project is one thing; whether the investment in the previous project can become the starting point for the next project is another.

As Zaniel calculated at the beginning, AI has already brought development costs down from 90 to 1, and the remaining 10 is the truly tough nut to crack.

If the processes that work today, the pitfalls encountered, and the experience accumulated can all be retained and turned into capabilities directly reusable for the next delivery, then there is a chance for that 10 to be gradually amortized away.

And Agnes Harness is walking exactly this road from 10 to 0.

Agnes Harness open-source address: https://github.com/AgnesAI-Labs/agnes-harness Agnes Sys1 1.0 technical report: https://github.com/AgnesAI-Labs/Agnes-Sys1-1.0-Technical-Report Agnes Harness technical report: https://github.com/AgnesAI-Labs/Agnes-Harness-Technical-Report-Nov-2026

  • Jensen Huang no longer silent: I am willing to take charge of Nvidia again! Board: Not considering it 2026-11-18
  • Vocational school student creates a viral "Harry Potter" hit! European and American stars and Hollywood producers collectively like it 2026-11-14
  • Pure code creates "Backrooms," switching 20 art styles in 116 seconds! Ernie's new model is having a blast 2026-11-11
  • Breaking! DeepSeek opens the full suite API 2026-11-14

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by 量子位
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗