william fischer (@williamfischer) header
william fischer (@williamfischer) profile picture

Exited Founder | CTO @ VONQ | Chair @ https://teapawt.net/t.co/7Gy4CJ9S7T | Board Advisor | Dad. Lapsed classical musician.

London, England
Joined February 2008
Enterprise AI Pilots: fail Long-Horizon agentic workflows: fail Billions and Billions spent to solve this. The solution. Enterprise software was built for humans. It needs to be built for Agents. Me explain. note on video: words are mine. not the voice. trust me, it's better that way.
1
1
1
5
535
inspired by grok bot, moves by @HubSpot , ANT:NASDAQ Can a company like @Workday can go from $10 billion to $200 billion. Services as a Service is the path. Just need the right harness.
34
Super interesting. We landed in a similar place. youtu.be/NFcX-KlVhtw?is=vfU5…
Check out Code Mode, a new, experimental way for Claude Code to work with tools. Claude writes scripts that chain your tools together in sophisticated ways, and it can answer you while long commands are still running. Opt in with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 code.claude.com/docs/en/code…
54
Thinking in Public | William Penn's Hat Aneel Bhusri | Co-founder & CEO, Workday "if people want to do headless, they work through our APIs" @bhalligan @dharmesh We're now the Customer Harness Company." Isabel Espinapina | VP of Product Development, ADP "helping clients accelerate the delivery of employee, manager and HR experiences through their preferred AI agents." Lived in philly in the late 80s.   For a big city, it lacked big skyscrapers. Unwritten rule... new buildings shouldn't exceed the height of the Hat on the statue of William Penn. One day a developer decided to ignore this and it was game on. Big Skyline glow-up. Each building, on its own, had a compelling ROI model. Problem was if everyone pursued the same strategy, it worked for no one. Which brings me, obviously, to Headless. For decades, big ATS and HCM companies built their businesses around proprietary applications, workflows, and user interfaces. But, users (people and increasingly AI Agents) are interested in a new way of interacting with a system of record, headless. Users want to use a harness like Claude Cowork and access their applications via MCP. Advances in frontier models and the introduction of MCP make this seem like an easy path. But if everyone is looking at the same solution, then  what is the differentiation? A better MCP server won't cut it: just a thin wrapper easily replicated. Better tools, better documentation?  Maybe, pricing? MCP makes legacy software accessible, but it doesn't make it good at supporting autonomous work.   Companies need to consider building their own harness, rebuilding their legacy tech stack, support alternative pricing models, expanding their offering (even across verticals/markets), solve for permissions and governance in a unique way, etc.  Need to be token efficient for agents, allow for performant long-horizon agentic workflows for the ROI case, and need to solve the human-in-the-loop UX in a human-friendly way. If they all use the same harness (hermes, codex, claude cowork), then they are all allowing the layer where the work happens and the business intelligence to move to that third party harness.   If there were just one player choosing this headless path, it just might work. As there are thousands (including frontier model providers) plus many ai-native startups, existing players can't just build another tall building.
3
1
172
Challenge to differentiate just on the MCP server layer. Easy to replicate. It's the back-end that needs to be optimized for agentic workflows. Current back-ends were built to support human UX. There's a better way to code enterprise software for Agent UX. Also, suggest SaaS will need to build their own harness. 3rd party harness can be a band-aid but further makes it difficult to differentiate and also allows the work performed on the product to be performed on someone else's platform. This also allows organization intelligence to move to the harness layer. The existing SaaS then just becomes an expensive DB.
I came to the conclusion that most SaaS should become an MCP server. we’re getting the economics of AI software backwards. A lot of AI startups are building their own chat interfaces, agents, and workflows, essentially reselling tokens from the big labs at low to negative gross margins. What if SaaS companies did the opposite? Take HR software. Instead of building an AI assistant, become the best MCP server for HR. Expose tools to: - Benchmark salaries against proprietary market data - Run compensation reviews and calculate pay equity - Model salary increases against a department’s budget - Recommend promotions based on company policies - Update employee compensation in the HRIS - Manage approvals, permissions, and audit trails Let ChatGPT, Claude, or Copilot handle the conversation, reasoning, and orchestration. Let the customer pay for the tokens. The SaaS company keeps the valuable stuff: proprietary data, integrations, business logic, permissions, compliance, and the system of record. You still charge a SaaS subscription, but you’re no longer paying to generate every answer. You can ship faster, with fewer engineers, and potentially restore software-like gross margins. The next generation of SaaS might not have much of a UI at all. Just an incredibly good MCP server.
3
1
4
118
Nice validation watching the HubSpot pivot. We landed on a "harness" operating system as the only way to manage complex, long-horizon tasks. Here's our HR Harness paired to our "agent friendly" enterprise application performing an "effective date" action
Notes From the HubSpot Refounding, Day 1: Yesterday (November 12th, 2026), Brian Halligan (@bhalligan) came back as CEO of HubSpot and we declared that it was the refounding of HubSpot. You can get to the original post by visiting refounding .com. Here are some thoughts/notes for those following along at home, that are curious: 1) Yesterday, we had an all-hands meeting (I've stopped calling it the "all-minds" meeting because that term never stuck). We did it live from Brian's place. I opened and then passed it to Brian who did the main presentation. I've seen Brian give a *lot* of presentations. He's very good at it. This was one of his better ones. I live-Slacked his presentation in the company Slack channel while he was giving his presentation a few feet away. It was fun -- and funny (I hope). There were hundreds of people posting to Slack live. It wasn't just me. People loved the presentation. Really positive response. The energy was palpable. HubSpot is ready for a refounding. I'm excited! 2) A common question we've gotten from folks in the community: "What does refounding mean? What will actually change?" We are reimagining, rethinking *everything*. The core question we're asking ourselves is: "Selves, if we were starting HubSpot today, knowing what we know, what would we do?" To start with, we're sharpening the product vision. We are leaning strongly into a headless version of HubSpot. This means that an increasing amount of HubSpot's capabilities will be consumable by AI agents and AI harnesses (like Claude and ChatGPT) via MCP. We're shifting significant resources towards that vision. It's how people increasingly want to use/access HubSpot and we hear that loud and clear. 3) After the presentation, Brian and I had a celebratory founders' dinner. Talked about a lot of things. But there's one overarching thing. We both agreed that we should BE BOLD. Take risks and don't hold back from necessary change and short-term sacrifice as long as we think it Solves For The Customer long-term. So, we are now on Day 2. Stay tuned for more updates. If you have questions, comments or feedback, just drop them in the comments. Cheers.
188
"Customer Harness Company" Which Applicant Tracking System will become the first Applicant Harness System notablackbox.ai has some ideas.
I'm back as CEO of @HubSpot today. @dharmesh and I are completely re-founding the company. Its day 1. We're now the Customer Harness Company. In the next few days, I'll post about why I'm coming back and what my plans are. LFG!
34
Notablackbox.ai gave their agents intelligence. Then they gave them computers, slack, phones, and inboxes. HR Harness + Agent friendly Application wut can't they do?
38
hr harness paired to optimized system of record in action. open to better ideas on making a headless SoR work, but pretty convinced this is the way.
33
Thinking in Public | if you understand just this one thing.... Flipping through a summary of the @AnthropicAI S-1.... Lots of interesting stuff, but so many words/numbers. incredible company, btw. my takeaway for #HRTech people see the value of Anthropic as their models (opus, fable, sonnet) smart investors will see the value in their Harness (i.e. claude code, cowork) from their S-1, "a model alone is not an agent" the model is the intelligence, but it's the harness that allows Claude to actually perform the work. investors look at the potential value of the "work" and see that Anthropic will capture a reasonable share of that value creation. this is the opportunity for HRtech. it's how a company will capture the $650B spend in this category. by capturing the value of the work the software automates. this happens in the Harness. if an ATS like @Workday allows a 3rd party harness like Claude to be where work happens in the ATS, they are giving all of the value creation to Claude. savvy investors are seeing that this dynamic will play out across all listed companies. that ai automation will happen, but it will happen via Claude. And, Claude will capture all of the value. Not a Black Box, Inc. didn't set out to be a harness company. it was just trying to solve for long horizon agentic tasks. it quickly became clear, the company needed its own harness.
72
Imagine a Claude Cowork built for an HR team. Comms, governance, multiplayer etc. Imagine if Claude also built a system of record. Imagine that both were built to work together. One stack. Can do all things HR. Here's GSD mode which asks for permission less often. Live run in second half of video. from not a black box dot ai
1
55
Sequoia just made a bold claim: They think the next $1T company will NOT be a software company. Their take? Services will be the new software. Because SaaS companies have spent 25 years fighting over a smaller budget in comparison. For every $1 companies spend on software → they spend about $6 on services. For reference: • A company might pay $10K a year for QuickBooks. • And $120K for the accountant who closes the books. AI has completely changed that. Now, a company can sell the finished WORK instead of the tool. Here's how the 3 models compare: TRADITIONAL SAAS • The customer buys access to software, and their own team runs it. • Growth is capped by product + distribution. TRADITIONAL SERVICES • The customer buys people's time + expertise • Growth is capped by hiring. SERVICE-AS-A-SOFTWARE • The customer buys a completed piece of work. • Software handles most of the delivery, with the provider's team on decisions + exceptions. • Growth is capped by how much of the work software can do reliably. Something to keep an eye on.
2
49
william fischer retweeted
Sequoia just made a bold claim: They think the next $1T company will NOT be a software company. Their take? Services will be the new software. Because SaaS companies have spent 25 years fighting over a smaller budget in comparison. For every $1 companies spend on software → they spend about $6 on services. For reference: • A company might pay $10K a year for QuickBooks. • And $120K for the accountant who closes the books. AI has completely changed that. Now, a company can sell the finished WORK instead of the tool. Here's how the 3 models compare: TRADITIONAL SAAS • The customer buys access to software, and their own team runs it. • Growth is capped by product + distribution. TRADITIONAL SERVICES • The customer buys people's time + expertise • Growth is capped by hiring. SERVICE-AS-A-SOFTWARE • The customer buys a completed piece of work. • Software handles most of the delivery, with the provider's team on decisions + exceptions. • Growth is capped by how much of the work software can do reliably. Something to keep an eye on.
90
56
29
736
133,325
if the models converge in performance, but we can tune our harness and application layer to get 35x improvements. The value creation opportunity sits with the harness and application layer, no?
26
when enterprise software goes headless and it will. then the best back-end wins. in headless world, agents are the users. enterprise software wasn't built for that. not a black box dot ai fixes this with Dual Representation (patent pending). a new way of writing enterprise software that is agent friendly.
1
58
Latest from notablackbox dot ai. We've just added agent access to cloud computers when using our vertical harness (built for HR). Greatly expands their capabilities and reduces latency and token costs. more here: AI Agents Are Getting Their Own Computers. HRTech Leading in Vertical Harness Development prlog.org/13176612-ai-agents…
1
44
if the enterprise software is moving headless. then the competition moves to the back-end. front-ends have multiple representations depending upon people UX back-ends also should have multiple representations to optimize for Agent UX Notablackbox.ai built 2 (filed a patent application for 1 of them) and are building a POC for a third. Best back-end wins.
26
Building the Machine that Builds the Machine | in Public as a quasi-hobby in early 2025, started a tiny ai lab, to solve for agentic long-horizon tasks. quickly figured out, that the challenge wasn't the models. it was "workday." we needed to rebuild "workday." so, we needed to develop a new way of building. video for deets. if, like me, you do 2x speed. volume down. unless speedcore is ur jam. turns out, not my jam. for the curious: we've got lots of agents building 24/7 every morning I get a slack message - see below (just longer) - with a summary of what they've been up to. Release Log — 2026-10-11TL;DR: A big hardening day (70 PRs across both repos) with two threads running together. Claw channels got their conversational core (an inbox every run reads, child runs per outside conversation, and a message tool over email), and the Office got leaner campaign and outputs views behind flags. Alongside that, a QA-sweep cleanup closed a long list of 500s,
1
71