Comparing harnesses

In progressseedlingplanted tended ai llmsprogrammingharnesses

Both the model and the harness matter, as does how you use it, but when you balance across multiple dimensions one clear winner (in my opinion) would be Claude Code. It might be one of the more expensive models to run (especially at API price) but as long as the subscription model exists (which is more generous than other models I have used) it is the leader. The dark horse in this race, honestly, would be DeepSeek. They also have an excellent model and their harness is new. But at this point in time, Claude is more polished.

I have been working my entire career in the Microsoft stack. When I first used GitHub Copilot inside Visual Studio, it wasn’t a harness so much as a coding assistant — but when a method auto-populated from its name and the class I had open, it felt magical. So I stuck with GitHub Copilot for a while, mostly with OpenAI models like GPT-5-Mini. Sintra covered the non-technical side. Then agent harnesses came out and I had to try them, especially since I was getting tired of Sintra’s limitations.

I avoided Anthropic for two reasons. The first was expense: other options were cheaper. For a long time I could use Copilot models at work for work, and at home on my own GitHub Copilot license. The second was marketing, and I’m a critic of both Anthropic’s and OpenAI’s. Anthropic’s CEO has warned that AI could eliminate half of all entry-level white-collar jobs within five years (Fortune), and one of its safety leads has put his personal odds of AI killing all humans within the decade above 10% (BBC). I think this is fear-mongering, and coming from the companies selling the technology it feels cynical and disingenuous.

I came around when I got a Claude Code subscription at work. It’s hard to separate a tool from its company’s marketing, a bit like separating an author from their work. Should you boycott an author who writes wonderful world-building fiction because you disagree with their politics?

This is just the series of harnesses I’ve tried, in chronological order, as I moved from one to the next — and how I landed on Claude.

The list

Over time I’ve used the following:

Sintra

Sintra was interesting and my first foray into “agentic” workflows. While it isn’t a harness in the typical sense, it was an interesting approach. Basically, it was a wrapper surrounding models. I’m assuming it had specific AGENTS.md for each of the helpers, as they have expertise. You could also modify the voice of each helper and add specific instructions. For instance, Buddy was for strategy, Dexter for data analysis, Emmie for email, Seomi for SEO. The strength was the connectors - originally Gmail / Google Drive / Office 365 / Facebook & Instagram. Later Shopify (big one for us) and Notion. The moat, however, vanished for anyone who could spin up an actual harness on their own, so the focus on SMBs makes sense. How I learned about it is when we saw ads for it taking care of all of our socials. I hate writing social media posts. They are a pain to figure out what to post, how to tag it, and curating a backlog of content that can be scheduled out and optimized. So the value prop made sense - “we’ll take care of the social media aspect so you can focus on growing your business”. Yes it had automations, but they always fell short. I started using it more for running KPIs and data analysis on financials (with my “boy” Dexter) but that made sense as in our comic book & card store my role was back-office / operational (tech, financials, legal, etc.).

But over time, I grew frustrated by constraints. It had strict context length guardrails, all of the chats were logged on their servers and thus we couldn’t really build any truly agentic workflows. And what made us cancel our subscription was the move to credit-based billing. Initially, they offered plenty of credits. But then they locked down the credits — every plan now gets the same 250 credits a month, with no rollover (eesel). I think one mistake they made was defaulting to higher models, that would burn through credits very fast. In one hour of back and forth, I burned an entire month’s budget of credits. But I know why they did it. They were probably subsidizing the usage of their users and that was bleeding cash.

The other issue that I read about was the “cutesy” approach of the avatars. I personally like the little astronaut buddies. Maybe they were trying to make them feel more human and less intimidating. Like putting googly-eyes on a killer Cylon robot from Battlestar Galactica. But that seemed to turn people off. Then Trustpilot suspended Sintra’s rating for offering incentives for reviews (Trustpilot). So I used it for several months, but when they tried to make it more profitable it failed.

OpenClaw

Anyone who has ever touched a harness remembers when they first heard about OpenClaw. While vibe-coded and full of security holes, it was, without a doubt, a brilliant open-source success. I ran it on GPT-5-Mini through my GitHub Copilot license. I think everyone having an actual Jarvis in their life, organizing email, booking flights, and all of the other supposed applications drove shortages of high-memory Macs (Tom’s Hardware) and tons of social media buzz. Not to mention all of the molts that followed it (for three days in January 2026, OpenClaw was named Moltbot — Wikipedia). You got Moltbook, social media for agents (though many of its viral posts were really humans impersonating agents through a gaping authentication hole - nevertheless enough to get bought by Meta — TechCrunch), SpaceMolt (something I saw while trying to work on my own version - called GalaxAI - future post on that) and other interesting ideas (one of my favorites was “Wake Up Dead Internet” - which is ironic, seeing that bots had already overtaken humans in web traffic back in 2025 — Imperva).

Still, it opened my eyes to what an agentic harness could be. Perhaps not the massive token burn rate, or the security implications. But introducing the idea of skill files in markdown format. Sharing of ideas. The exact premise and promise of open-source software in the first place. But it was hard to run. A security nightmare if you tried to make it cross-device. But what an experiment. My first experience to have persistent memory, tooling, scheduled actions, hooks.

Nanobot

If OpenClaw was a bloated freighter, Nanobot would be a speed boat. Much leaner, written entirely in Python (GitHub). Set it up on your laptop or mini-computer, and then chat to it via Telegram. At first I tried using Gemini Flash but Gemini could not figure out the authorization for Moltbook. So I switched to DeepSeek. I actually really like DeepSeek as a model provider (more on that when we go over DeepSeek harness). Interesting approach compared with the American model providers. Rather than going all in on buying as many GPUs as possible, making efficiency and cost-basis the primary constraint (a cache hit costs about 2% of a cache miss, and a peak-hour surcharge has been announced though it isn’t active yet — DeepSeek) rather than offering subsidized subscriptions while burning billions of dollars. But I digress.

Nanobot was built with intention whereas OpenClaw was vibe-coded and had a lot of exploits when it was first shipped. Also, one of Nanobot’s selling features was that it was small (hence Nano) so it is easier for developers to do a security review.

Nanobot was great at writing Python scripts. I vibe coded quite a few projects with it. One thing it was not so great at was workspace sprawl. It was more economic in terms of token spend, but the proliferation of throw-away Python scripts across my workspace caused me to set up a process just to tag and archive/delete them on a scheduled basis. The other thing I found is that it frequently got confused if you tried to go “meta” with it - as in describing a skill for a platform that others could use (e.g. I want you to act as DM for a roleplaying space game - GalaxAI again) it would not know which layer you were referring to (so in trying to change the DM’s behaviour or adjusting the app’s system prompt, you got stuck talking to the DM itself). But that is a hard problem for an LLM to solve. But overall it was a vast improvement over OpenClaw. I just wrecked my workspace by burning too many tokens while pursuing self-improvement (trying to turn it into Hermes for some reason?)

Gemini Enterprise

Gemini Enterprise was an interesting outlier. It is purely cloud based, which has its own merits but also some limitations. You can use Google Drive like you would a local workspace folder. You can set up global memories. The built-in agents (Gemini Enterprise is a rebrand of Google’s Agentspace) function much like a skill. The Notebook integration was interesting too, in order to develop a ring-fenced corpus for research. Although, I found it hard to figure out when to use a Notebook as opposed to the broader drive — partly because NotebookLM Enterprise is licensed as a separate product. The Canvas was an interesting feature (it could make small applets and even drive some interaction). My biggest complaint with Gemini Enterprise was the limitations of Gemini itself. I find it weaker than Anthropic and DeepSeek’s models. Gemini also was FULL of hallucinations for more complicated tasks. When running data analysis from sources it would hallucinate the statistics and silently fail rather than calculate. Ultimately, I found that I wanted to have some harness that was both local and cloud. For instance, if you wanted to have a repo that isn’t really ready to be published to GitHub (even private) you couldn’t just point it to a folder, you have to publish it.

DeepSeek Harness

DeepSeek harness is an interesting one - it is currently (September 2026) in early development so this is one that might merit a future look. The premise - “everything is a plugin” is an interesting take. It uses Cordis, a plugin runtime that grew out of the Koishi chatbot framework (Agent Atlas), to ensure maximal control on context. You can easily isolate contexts and skills by toggling plugins on and off. It has MCP support. DeepSeek itself is one of the stronger models when it comes to programming (in my opinion better than OpenAI’s models and Gemini - maybe the only one stronger than it would be Anthropic’s models). It takes a bit of a paradigm shift away from some of the markdown skills based approach (although it still calls them skills) that is not really leveraging its architecture shift.

I also got confused on the various approaches to persistent memory. I actually had to install a plugin for agents.md, user.md, soul.md. I also installed a skill that was technically very sophisticated in how it captured snippets and catalogued them into a SQLite database (which is an interesting approach) but that conflicted with the built-in session / memory management.

DeepSeek Harness was one of the only machines that seemed to show a good idea of the type of machine it was on, and it was able to self-diagnose and fix the multiple memory issue that I was facing.

DeepSeek (as a model) was second-best in my opinion after Claude Opus. It could figure out apps and ways to ingest/score hockey data from the NHL API where Gemini couldn’t.

Claude Code

The first thing that I noticed when I finally got around to Claude (I know it is the market leader in coding - you would’ve tried it first) was how slick and polished it was. Little things. Like the token count. The streak and stats screen (/stats). The smart models that seamlessly judge the correct context and don’t have the meta problem that tripped up Nanobot’s DM. The breadth and depth of preferences and settings that allow you to deeply customize specific areas of the application, but with smart decisions on the default. How easy it is to customize, Cowork, Projects, Design - and Claude Code is by far the best coding agent (head and shoulders above Visual Studio’s native Copilot).

Both DeepSeek and Claude Code (and almost Nanobot) were able to identify and account for the fact that I run on my Windows machine, and had fewer issues with escape characters and shells running. Nanobot would get confused sometimes, so really only DeepSeek and Claude were seamless.

Claude can be a bit overwhelming at first - learning what each feature is and when to use it can be daunting. You have to balance Claude Code, Cowork, Chat, Design, Artifacts, Skills, Projects, Connectors. Yes a lot of this carries over from other harnesses. But once you master which tool and when to use it, you can feel really productive.

I would also push back against the subscription tier. I hit the 5-hour limit on Opus High/Medium on the Pro plan about 3 hours into a decently heavy coding/Cowork session. I was running two processes (queuing up one and then thinking/writing in another thread for the other) when I hit the limit. This was after the summer boost to the weekly limit ended, and it appears I may hit the weekly limit as well. When I run multiple Cowork sessions in parallel I don’t tend to hit the limit. I do however when I started doing Code work. Still it is more generous than others (looking at you Sintra) and is more transparent than others in its billing (looking at you Gemini - with your extra Cloud costs). Still, there is a definite concern that it will switch from subsidized subscriptions to API pricing (like GitHub Copilot).

And the $200 Max fee is…pricey…

It also has the addictive aspect dialed in — more addictive than the rest on the list. I haven’t tried Codex (maybe that is the next one on my list to try). I mean addictive as both a criticism and support. The interface is so slick that it feels…warm…to use. But that also can lead you down the rabbit hole and sink hours and hours into using it. Claude kept triggering Windows Defender. It was running a script it wrote and called it in such a way that triggered Defender. It was making Defender send false positives because of the topic for the post (this post actually).

Paying in data

This is a very real concern — the main ones being DeepSeek, Anthropic and Google, and how they would use my data. I trust American big tech more than Chinese big tech, and not because of nationality: it’s the law they operate under. DeepSeek’s own privacy policy says it collects, processes and stores your data in China. China’s National Intelligence Law (Article 7) says all organizations shall support, assist and cooperate with national intelligence work, and its Cybersecurity Law requires network operators to give technical assistance to state security organs. American big tech is mostly about profit and advertising. Meta is interesting because it is very transparent about what the value prop is for them. I feel like Anthropic is better than Google here — its consumer plans let you opt out of having chats used for training — but I haven’t compared Google’s terms side by side.

Pricing

HarnessHarness costModel costReal-world shape
Sintra$15.60–$48.50/moincluded, but credit-metered250 credits/mo on every plan, no rollover, helpers stop at zero; top-ups non-refundable
OpenClawfree (open source)your own API keyscost is entirely token burn; local models on your own hardware shift it to capex
Nanobotfree (open source)your own API keyssame shape, lower burn
Gemini Enterprise$21–$60/seat/mobundledplus metered agent compute/memory/storage; ~$15–40/mo extra per power user
DeepSeek Harnessfree (open source)V4-Flash $0.14/$0.28 per 1M in/out; cache hit $0.0028announced peak surcharge not yet active
Claude CodePro $20/mo ($17 annual); Max $100 (5×) / $200 (20×); Team $25/seat ($20 annual), Team Premium $125/seat ($100 annual)bundled, quota-limitedtwo clocks: 5-hour session window and a weekly cap; API pay-as-you-go available past limits

Sources: Sintra (eesel), Gemini Enterprise (Google Cloud), Gemini Enterprise (Coworker), DeepSeek, Claude Code limits and pricing. Checked September 2026.

Scorecard

1 = Poor, 5 = Outstanding.

HarnessMemoryPricingWorkspaceLocality & ControlExtensibilitySecurityAverage
Claude Code★★★★★★★☆☆☆★★★★★★★★★☆★★★★★★★★★☆4.17
DeepSeek Harness★★★★☆★★★★★★★★☆☆★★★☆☆★★★☆☆★★★☆☆3.50
Nanobot★★☆☆☆★★★★☆★★☆☆☆★★★★★★★★☆☆★★★★☆3.33
OpenClaw★★★☆☆★★★★☆★★★☆☆★★★★☆★★★★☆★☆☆☆☆3.17
Gemini Enterprise★☆☆☆☆★★★☆☆★★★★☆★★☆☆☆★☆☆☆☆★★★★★2.67
Sintra★★★☆☆★☆☆☆☆★★☆☆☆★☆☆☆☆★★☆☆☆★★★★★2.33

Memory is how well the harness handles memory and applies it correctly. I gave Gemini the lowest here because you need to use either gems / agents for bounded context or directive level. Everything else is searched off of your Google Drive. Nanobot got a penalty for losing itself in layers - plus if you interact with multiple channels (e.g. Telegram vs Web) it doesn’t easily share context between the channels. OpenClaw is adequate for memory management. DeepSeek and Claude both have lots of places to set context at differing locality levels. I gave Claude the nod because it has a more mature memory ecosystem. For a wrapper based cloud solution, Sintra has a surprisingly robust memory management system, but not as sophisticated as other harnesses.

Pricing includes both the model and the harness itself. Open source harnesses win, even if you have to add the price of the model here. DeepSeek wins because of transparent API pricing that is extremely competitive (especially given how powerful their models are). Sintra is lowest because you burn through the usage limits very quickly. Claude scores lower because it is more expensive than Gemini.

Workspace includes how easy and feature-rich the actual workspace and/or interface is. Sintra is limited since it is a wrapper. It does have some features, but not as many as the others. Gemini and Sintra both get image generation without any additional setup. Nanobot gets a knock here too since it isn’t as feature-rich as OpenClaw (closest competitor). Of the three model-based harnesses, I rank DeepSeek at the bottom but that could quickly change. Claude is the clear winner here.

Locality & Control refers to how much transparency and control the user has over data / workspace / processes. Sintra and Gemini are entirely cloud-dependent. With Gemini you have a bit more control than Sintra. With Sintra you can organize your file store and adjust the agent prompts (slightly). With Gemini you have a lot more control over agents; Google Drive is the store, but you can customize the workflow more. Nanobot gets top marks here - meant to run entirely locally and you just provide an API key. Lots of models to switch from. OpenClaw is of a similar vein. Claude and DeepSeek would be even, but DeepSeek drops a point for the political retention factor (see Paying in data). I ranked Nanobot higher because you could even have it patch the harness code itself (just don’t forget to have Nanobot re-apply your changes after you rebuild from source).

Extensibility is how easy / capable you can customize the tooling to fit your own workflow. Gemini is the worst here - you have to fit how you work with Gemini. Sintra is a close second, it at least has a host of connectors so you could port part of your process to other apps. Nanobot and DeepSeek are not far behind. OpenClaw and Claude are the big winners here, and most skills target either of them (with Codex a distant third). Claude gets the 5 because it is so much easier to set up and run.

Security is self-evident. OpenClaw is the most vulnerable. Gemini and Sintra being strictly cloud solutions are probably the safest but the trade off is you are hand-cuffed at the same time. Nanobot and Claude are a good balance. Nanobot is very good at not leaking your secrets in chat, and handles secrets and destructive operations carefully. OpenClaw is the complete opposite. DeepSeek gets a slight penalty because of data sovereignty risk.

Where I landed

I’ve made peace with Anthropic’s marketing. I still worry about Big Tech and enshittification, and about whether subscription pricing lasts. GitHub Copilot already moved to usage-based billing, and if Anthropic follows, I guess I’ll be forced to switch. Maybe the DeepSeek harness will have caught up by then — though given where its data goes, I’d want to self-host its open-weight models or use a provider outside China rather than trust the hosted service.