Loading articles…
Field notes for building a business that can move without you.
We write from the intersection of operating systems, software, sales, marketing, and leadership—for owners building a more capable company.
Engineering deep dive
The intelligence inside our agentic harness—and how it gets smarter, not dumber, over time
Small businesses are being told to wire AI agents into a dozen apps. We built the opposite: one opinionated operating system, and a harness that lets the agent grow more capable without growing more confused.
- Author
Alexander Ford- Published
- Reading time
- 20 min read
Contents
In brief
- For a decade, small businesses have been told to run on a stack of single-purpose apps stitched together with integrations. Now they are told to do the same with AI agents. The evidence says both approaches mostly stall.
- The failure is structural. Every tool you connect to a general-purpose agent spends some of its limited attention, and a patchwork of apps gives it no shared meaning to reason over.
- Telos takes the other path: one opinionated model of how a business runs, and a harness—the software around the language model—built so capability can grow without clutter.
- Sapience, the agent inside Telos, can use 145 tools but keeps 27 in view at a time, cutting the tool descriptions it reads on every step by about 81%.
- Every tool is defined once and shared by every agent we run, so each improvement lands everywhere at once. That is why adding capability has made our agents sharper, not duller.
1. Baling wire and duct tape
If you run a small business, you have heard the advice for years. Buy the best app for each job—one for customers, one for projects, one for invoices, one for scheduling—and connect them. When the connections break, add another integration. When nobody can find anything, add a dashboard.
The results are well documented. In one survey of U.S. businesses with 10 to 99 employees, the typical company ran about ten separate business tools, spent roughly $10,000 a month on software, and wasted about $3,000 of it on apps nobody used. Teams reported spending around 25 hours a week re-entering or reconciling data between those apps, and 91% said the manual reconciliation lowered their productivity.[1] (That survey comes from a software vendor, so read it as directionally right rather than exact.) Researchers watching office workers found them switching between applications about 1,200 times a day, losing close to four hours a week just reorienting.[2] And the businesses that grow tend to be the ones that escaped the patchwork: 66% of growing small businesses report integrated systems, against 32% of declining ones.[3]
Now the same advice has arrived for artificial intelligence. Automation platforms advertise AI agents that work across thousands of apps. Chat assistants plug into your tools through connectors. A growing chorus suggests you skip the software altogether and have an AI build you your own CRM over a weekend. It is the old strategy with a new engine: you are still the systems integrator, only now the thing you are integrating can also make things up.
It is not working for most small businesses, and the numbers say so plainly:
- Among the growth-minded owners in Goldman Sachs' 10,000 Small Businesses program, 76% use AI, but only 14% have fully integrated it into their core operations.[4]
- McKinsey's 2026 survey found large organizations scaling AI agents jumped from 27% to 40% in a year. Smaller organizations stayed flat at 22%.[5]
- U.S. Census data shows AI use among firms with fewer than 20 employees did not change significantly from late 2025 to spring 2026.[6]
- MIT's widely cited (and widely debated) study of enterprise AI reported that about 95% of organizations saw no measurable return, and named the cause: “brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.”[7]
- Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027.[8]
None of this is because owners are doing it wrong. It is because the approach asks a ten-person company to do what large enterprises with IT departments struggle to do: make a dozen systems that were never designed to agree behave like one business.
2. Why bolting tools onto an agent makes it dumber
To see why, it helps to know what an AI agent actually is. At its core is a language model—the part that reads and writes. Around it sits software that gives the model tools: the ability to look up a customer, create a task, send an email. On every step the model reads everything placed in front of it—its instructions, the conversation so far, the results of earlier lookups, and a written description of every tool it might use—then decides what to do next. That working space is called the context window.
Context windows are large now, but attention inside them is not free. Research across eighteen leading models found that performance degrades as input grows, even when the task itself stays equally hard.[9] In another benchmark, eleven of thirteen models that advertise very long context windows fell below half of their short-context accuracy by the time the input reached 32,000 tokens (about 24,000 words).[10] Models also reliably lose track of information buried in the middle of long inputs.[11]
Tools make this worse, because their descriptions ride along on every step whether they are needed or not. Anthropic measured one popular connector's 35 tools at roughly 26,000 tokens, and reported setups where tool definitions alone consumed 134,000 tokens.[12] When researchers grew a tool catalog from 8,000 to 120,000 tokens, models got between 7% and 85% worse at using them.[13] In another study, loading every tool at once left the model choosing the right one only 13.6% of the time, against 43.1% when it was shown only the relevant few.[14] OpenAI's own guidance is to keep fewer than twenty tools available at the start of a turn.[15]
“Have you heard stories like: the more tools you give an agent, or the more MCP servers you hook up to it, the agent gets dumber? People try to give an agent 500 tools and the agent becomes stupid.”
— Alexander Ford, on a client call, August 2026
There is a second, quieter failure. A patchwork of apps has no shared meaning. The “customer” in your CRM, the “client” in your invoicing tool and the “account” in your project software may be the same company, or may not. People resolve that from memory. An agent has to guess, and it guesses fluently. Research on business data bears this out: when language models answer questions against a well-defined semantic model of the data instead of raw tables, accuracy rises sharply—in one 2026 benchmark from 90.0% to 98.2%—and the failures change character. As the authors put it: “With text-to-SQL, failure looks like a plausible but incorrect answer. With the Semantic Layer, failure looks like an error message.”[16],[17]
The more a general-purpose agent is connected to, the less of its attention is left for understanding any one thing it is connected to.
3. The other path: an opinionated system with a harness built for it
We build Telos, the software inside the Accelerator OS, and we took the opposite approach. Instead of connecting an agent to everything, we gave it one coherent place to work, and invested heavily in the part of the system most people never see: the harness.
The model is the brain. The harness is the body.
A harness is everything around the language model that turns it into something useful: its identity and rules, the tools it can use, how those tools are described, what it remembers, how its working space is kept clean, and the guardrails that stop it doing harm. The model decides; the harness decides what the model sees when it decides, and what its decisions are allowed to do. The best harnesses, in our experience and Anthropic's, are built from simple, well-understood patterns rather than elaborate frameworks, with most of the effort spent on the tools themselves.[18]
| Layer | Role | Parts |
|---|---|---|
| Model | The brain. Any frontier model, chosen per conversation. | Reasoning; Language; One model for the whole turn |
| Harness | The body. Decides what the model sees, can do, and must never do. | Identity and rules; Core tools + load on demand; One shared tool registry; Context hygiene; Memory; Durable runtime and guards |
| Telos | The world it operates in. One agreed shape for a business. | Strategic plan; Customers and deals; Projects and tickets; Knowledge and its graph; Activity timeline |
Figure 2. The model reasons; the harness curates what it sees and governs what it does; Telos supplies a business with a known shape to act on.
Because the harness carries the intelligence of the system, the model underneath is replaceable. Telos works with a range of frontier models, and when a better one arrives, every capability we have built carries over to it on day one.
An opinionated model of the business
The first design decision is the least technical. Telos has an opinion about how a business is shaped. Strategy follows a one-page strategic plan: a foundation (purpose, vision, a long-range goal), multi-year horizons, strategic thrusts, an annual plan, initiatives, and quarterly priorities we call rocks. Customers, deals and offers live in one place. Work is projects and tickets with owners, checklists and dependencies. Knowledge lives in a searchable library connected by a graph of typed relationships. Every call, email, meeting and change lands on one activity timeline.
For a person, that structure removes guesswork. For an agent, it is the difference between reading a map and reading a pile of receipts. When Sapience is asked how the quarter is going, it does not scrape three apps and hope the names line up. It reads rocks that are typed, owned, dated and linked to the initiatives they serve—and its strategy tool can point out what the plan is missing, such as a rock no initiative depends on or a thrust with no initiatives beneath it.
Picture a fireplace retailer on a busy Saturday. A couple walks in looking at a gas insert. Afterward the salesperson types one sentence: “Log the Hendersons from this morning, they want a quote on the gas insert, and remind me to follow up Thursday.” Because Telos already knows what a contact, a deal and a task are and how they connect, the agent finds or creates the contact, opens the deal in the right pipeline, sets the task, and records what happened—on the same records the rest of the team works from. Ask it next week to draft a follow-up, and it writes from what actually happened, not from what it half-remembers.
Many tools, few in view
The second decision is how capability is exposed. Sapience can use 145 tools today, grouped into 11 categories—strategy, sales, operations, knowledge, communication and so on. But on any given step it sees only 27: the readers and finders that almost every conversation needs, plus a tool for loading more. When a request calls for something else—updating a strategic plan, merging duplicate contacts, proposing meeting times—the agent loads that category, uses it, and keeps it for the rest of the conversation.
| Tool definitions | Tokens per step |
|---|---|
| Every granted tool, described up front (145) | ≈66,700 tokens |
| The always-loaded core Sapience actually carries (27) | ≈12,600 tokens |
Figure 1. Tokens of tool definitions the model reads on every step. Loading tools on demand keeps 81% of that budget free for the business itself. Estimated from the production tool catalog on September 23, 2026.
Think of a good employee handbook. It opens with a table of contents, not with every policy printed on the first page. You read the chapter you need when you need it. Anthropic calls this progressive disclosure, and its own measurements show why it matters: letting the model search for tools instead of carrying them all raised tool-use accuracy on its evaluation from 49% to 74% for one model, while cutting token use by 85%.[12],[19]
The same principle applies to the instructions that come with tools. Each category carries a short operating protocol—how to handle duplicates, when to confirm before acting, which record is the source of truth—and those protocols load only with the category they govern. The agent reads the rules for booking meetings when it is booking a meeting, not while it is summarizing a sales call.
One definition, every agent
The third decision is where most of the compounding comes from. Every tool in our system is defined exactly once, in a single registry: its name, its inputs, its description, the category it belongs to, and which agents may use it. Sapience, our own internal agent (Xavier), and the connections our engineers use all read from that one registry. Sapience does not define a single tool of its own; an automated check fails the build if anyone tries.
That matters because tool descriptions are not documentation. They are instructions to the model. We write them under a standard we hold ourselves to: the reader is a model under token pressure, every word costs attention, and the author is programming the agent, not describing a product. In our experience, ten or twenty well-chosen words in a tool description can completely change how an agent behaves. Anthropic reports the same: in its own work, it spent more time optimizing tools than the overall prompt.[18],[20]
So when we sharpen one description—because an agent reached for the wrong tool, or skipped a step—that improvement reaches every agent the tool is granted to at once. We run our own company on the same harness. Every lesson our team learns operating Measurable Genius becomes a better tool definition, and every business on Telos inherits it.
Keeping the working space clean
Long conversations are where most agents quietly degrade: each lookup leaves its full results behind, and after an hour of real work the model is wading through pages of stale output. Our harness treats context as an asset with an opportunity cost.[21]
- Old results are folded away, not lost. When a conversation is replayed to the model, only the most recent round of tool results is kept in full. Older ones are replaced with a one-line note saying what they were and how to fetch them again if needed.
- Long conversations are compacted. As a conversation approaches the model's limit, earlier stretches are summarized while the most recent tool exchanges stay word-for-word. The complete transcript is always kept; only the model's working copy is condensed.
- The stable parts are cached. The agent's identity and its core tools are kept byte-for-byte identical from step to step, so the model provider can reuse them instead of reprocessing them. That makes each step faster and cheaper.
- A map, not a dump. At the start of a conversation the agent is told what exists in the workspace—how many records, contacts, companies and deals—not the contents. It retrieves the contents when the question calls for them.
The effect is measurable. Over the last 30 days, the busiest step of the average Sapience run used about 70,800 tokens—roughly 7% of the context available to it. The agent works with room to think.
Look it up; don't remember it
A general-purpose assistant answers from what it absorbed during training and whatever you pasted into the chat. Sapience is built to look things up. It answers from the records in your workspace, and when a search comes back empty it says so rather than filling the gap.
“It doesn't depend on memory. It will always look up the knowledge, it will always look up the specifics, and it will always use exactly the information in the system—never acting from whatever the deal was I did last week.”
— Alexander Ford
The same discipline governs what it does. It may only say it created, changed, searched for or sent something if the corresponding action ran in that conversation and succeeded. A failed action is reported as a failure, not smoothed over. Its operating rules require an explicit yes before anything destructive—deleting or merging records—after the agent has restated exactly what will happen. And when it sends an email or text message, it sends it one-to-one, from the account of the person who asked.
Memory that accumulates in structure
An agent that looks everything up still needs to get to know you. Sapience keeps several kinds of memory, each with a clear job: short operational notes about how your business works, a searchable record of past conversations, profiles of the people it works with, and the knowledge library itself, where plans, processes and decisions are written down and linked. Recorded calls and videos are transcribed and made searchable the same way.
The important word is structure. Most memory is not loaded automatically; the agent retrieves what is relevant when it judges it relevant. Nothing is stuffed into the context just in case. So as your workspace grows, the agent has more to draw on without carrying more on every step. Every plan you write, every call you record and every decision you capture makes the next answer better grounded.
Built to finish the job
Real work takes more than one step. Sapience runs on a durable execution engine we built called Praxis, which saves the state of a run after every step. If a server restarts in the middle of a task, the run picks up where it left off instead of starting over or silently failing. For larger jobs, the agent can hand pieces to up to four focused sub-agents working in parallel, each with its own clean working space, and gather their findings. Guards watch for the ways agents go wrong—repeating the same failing action, or writing in circles—and stop the loop so the agent explains the dead end instead of spinning. You can redirect a run while it is working, or stop it outright.
4. Why it gets smarter, not dumber
Put those decisions together and three loops emerge that run in the opposite direction from a bolted-together stack.
- Capability grows without clutter. On July 1, 2026, our shared catalog held 147 tools. Today it holds 211. The always-loaded core has stayed small throughout, so each new tool adds a capability without adding a cost to every conversation that doesn't need it.
- Knowledge accumulates in a shape the agent understands. Because the data model is fixed, what you add this month is as legible to the agent as what you added last year. Nothing needs re-mapping.
- Every improvement is shared. A better tool description, a new guardrail, a smarter compaction rule—each is made once and reaches every agent and every workspace.
We hold ourselves to measuring this rather than asserting it. On Xavier, our internal agent running on the same harness, we track first-answer resolution: the share of answers people accepted without correcting, re-asking or rejecting them. Over the last 30 days it stood at 97.7%, from 1,560 answers labelled since measurement began in August. Inside Telos, Sapience has completed more than 10,555 runs across 58 businesses since April.
| Measurement | Value |
|---|---|
| tools available to Sapience | 145 |
| kept in view on any one step | 27 |
| fewer tool-description tokens per step | 81% |
| first-answer resolution on our internal agent | 97.7% |
5. Objections we hear
“Won't all those tools bloat the agent's context?”
They would if they were all loaded at once. That is the failure we designed against. Describing every tool Sapience can use would take about 66,700 tokens on every step; the core set it actually carries takes about 12,600. The rest arrive only when a request needs them.
“Why not just connect ChatGPT or Claude to the apps I already use?”
For answering questions about one app at a time, that can work well. Running a business is different. It needs the agent to understand how a deal, a project, a plan and an invoice relate, to act on them under rules your team agrees with, and to leave a record anyone can check. A general assistant wired into separate apps inherits every inconsistency between them, and you remain the person who reconciles them.
“Isn't ‘opinionated’ another word for rigid?”
The opinion is about shape, not content. Telos has a view on what a strategic plan, a deal or a project is; your plan, your deals and your projects are entirely yours. That shared shape is exactly what lets people and agents understand each other's work without a translation layer.
“What happens when a better model comes out?”
It makes Telos better. The harness is independent of any one model provider, so a stronger model inherits every tool, rule, memory and guardrail we have built on the day we adopt it.
“Will it do things without asking me?”
It acts on your behalf, in the open. Everything it creates or changes lands on the same records you work from, where you can see exactly what it did. It confirms before anything destructive, reports failures as failures, and sends messages only one-to-one from the account of the person who asked.
“Is my data mixed with anyone else's?”
No. Each business has its own workspace. Every action the agent takes is scoped to that workspace and checked against what that workspace is allowed to do before it runs.
6. The harness is only as good as the business underneath it
The lesson of the last decade of small-business software is that tools do not organize a business. People do, with a shared picture of where the business is going, who owns what, and what has already been decided. An agent makes that shared picture far more valuable, because it can read all of it at once and act on it. But it cannot invent the picture.
That is why the Accelerator OS combines Telos with hands-on implementation: we install the structure alongside your team, and the agent grows sharper as the structure fills in.
Methods
Internal figures were taken on September 23, 2026. Tool counts and grants come from the production tool registry. Token figures sum each tool's name, description, input schema and usage protocol, estimated at four characters per token. Run counts and context use come from the Telos run log; context use is the mean peak input per run over the prior 30 days divided by the mean context window of the models used. First-answer resolution is measured on Xavier, our internal agent, from how people responded to each answer; it has not yet been measured inside Telos. The July catalog count is as stated on a recorded call that day.
References
- Intuit QuickBooks. 2024 Business Solutions Survey. Intuit, August 2024. 630 U.S. owners and executives at businesses with 10–99 employees. A vendor survey.
- R. Murty, S. Dadlani, R. B. Das. How Much Time and Energy Do We Waste Toggling Between Applications?. Harvard Business Review, August 29, 2022.
- Salesforce. Small and Medium Business Trends Report, 6th edition. Salesforce, 2025. 3,350 leaders of businesses with up to 200 employees.
- Goldman Sachs 10,000 Small Businesses Voices. Small Businesses Embrace AI — But Need Training and Support to Fully Harness It. Goldman Sachs, 2026. 1,256 program participants, surveyed January–February 2026.
- McKinsey & Company. The state of AI in 2026: On the road to ROI. McKinsey & Company, August 2026.
- K. Grundy, C. Breaux, S. Khatiwoda. Large Firms With at Least 20 Employees Biggest AI Users. U.S. Census Bureau, May 26, 2026.
- A. Challapally, C. Pease, R. Raskar, P. Chari. The GenAI Divide: State of AI in Business 2025. MIT NANDA, July 2025. Preliminary findings; its methodology has been publicly debated.
- Gartner. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. Gartner, June 25, 2025.
- K. Hong, A. Troynikov, J. Huber. Context Rot: How Increasing Input Tokens Impacts LLM Performance. Chroma, July 14, 2025.
- A. Modarressi et al.. NoLiMa: Long-Context Evaluation Beyond Literal Matching. ICML 2025, February 2025.
- N. F. Liu et al.. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the ACL, 2023.
- Anthropic. Introducing advanced tool use on the Claude Developer Platform. Anthropic Engineering, November 24, 2025.
- K. Kate et al.. LongFuncEval: Measuring the effectiveness of long context models for function calling. arXiv, April 2025.
- T. Gan, Q. Sun. RAG-MCP: Mitigating Prompt Bloat in LLM Tool Selection via Retrieval-Augmented Generation. arXiv, May 2025.
- OpenAI. Function calling. OpenAI API documentation, accessed September 2026.
- B. Ganz, T. Perigaud. Semantic Layer vs. Text-to-SQL: 2026 Benchmark Update. dbt Labs, April 7, 2026.
- J. Sequeda, D. Allemang, B. Jacob. A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model's Accuracy for Question Answering on Enterprise SQL Databases. arXiv, November 2023.
- Anthropic. Building effective agents. Anthropic Engineering, December 19, 2024.
- Anthropic. Equipping agents for the real world with Agent Skills. Anthropic Engineering, October 16, 2025.
- Anthropic. Writing effective tools for agents — with agents. Anthropic Engineering, September 11, 2025.
- Anthropic. Effective context engineering for AI agents. Anthropic Engineering, September 29, 2025.
See what this looks like inside your business.
Talk with us about the Accelerator OS and whether it fits the way your team works today.
