On this page
- Key Takeaways
- The Plumbing Became a Standard Before Most Companies Noticed
- Skills Turn Know-How Into Something You Can Version
- Agents Are Moving Off the Laptop and Into the Background
- The Digital Workforce Is a Management Problem Before It Is a Tech Problem
- Workflow Logic Gets Cheaper While Context Gets More Valuable
- Five Bets for 2027, and What I'd Do About Each
- Trust Is the Part Nobody Has Solved
- The Workforce Gets Built One Workflow at a Time
- FAQ
Key Takeaways
- The agent stack standardized fast. MCP, Agent Skills and a hosted agent runtime all arrived in under two years. The plumbing question is mostly answered.
- Skills are the sleeper trend. A folder with a text file in it turns how your team works into something an agent can run, version and improve.
- Agents are leaving the chat window. In 2027 more of them run on schedules and in the background, which makes ownership a real question.
- The "digital workforce" is a management problem first. Agents need a manager, a scope and a review cycle, the same as a new hire.
- Context is where the advantage goes. When everyone has the same model and the same protocols, the company with the best-written workflows and the cleanest data wins.
The most important agent release of the past year was a folder with a text file in it.
That's how Anthropic describes Agent Skills: a directory with a SKILL.md file, plus whatever scripts and reference docs the task needs. No new model. No new interface. And I think it matters more for how companies will use Claude in 2027 than any benchmark jump this year.
Here's why. For the last two years, most of the conversation about agents has been about capability. Can the model plan? Can it use tools? Can it run for an hour without going off the rails? Those questions are mostly settled for everyday business work. The question that's left is much more boring and much more important: does the agent know how your company does things?
The industry has spent the past 18 months building the answer to that question into standards. The Model Context Protocol gave agents a common way to reach your systems. Skills gave them a common way to carry your know-how. And now hosted runtimes give them a place to run when nobody is watching.
Gartner expected this to move fast. In 2025 it predicted 40% of enterprise apps would feature task-specific AI agents by the end of 2026, up from less than 5%. Whether or not that exact number lands, the direction is right. So here's what I expect custom Claude agents to look like in 2027, and what I'd do about each trend now.
The Plumbing Became a Standard Before Most Companies Noticed
A year and a half ago, connecting an AI assistant to your CRM meant a custom integration project. Somebody wrote glue code, and that glue code broke every time a vendor changed an API.
MCP changed the math. Anthropic released it in November 2024 as an open protocol. By December 2025 it had more than 10,000 active public servers and 97M+ monthly SDK downloads, and Anthropic handed it to a new Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI. ChatGPT, Gemini, Microsoft Copilot, Cursor and VS Code all support it.
That last part is the trend. When competitors adopt the same protocol, it turns into infrastructure. Nobody argues about which electrical outlet to use anymore.
The complication: a standard connector is only as safe as the permissions behind it. An MCP server that can read your whole data warehouse will happily do so for whoever asks. The protocol solved the "how do I connect" problem. It did not solve "who should be allowed to see what."
Treat connectors as shared infrastructure
In 2027 the smart move is to build each connector once, scope it by role, and let every agent reuse it. If three teams each build their own Salesforce connector, you have three security reviews, three sets of credentials and three things that break. One well-permissioned connector serving ten agents is a much better position.
Skills Turn Know-How Into Something You Can Version
Agents have always been good at general work and bad at your work. Ask Claude to draft an investment memo and you get a solid generic memo. Ask it to draft your firm's memo, with your section order, your risk language and your house view on add-backs, and it has nothing to go on.
Skills close that gap. The design is simple. At startup the agent loads only each skill's name and description. When a task matches, it reads the full instructions. If it needs the bundled scripts or reference files, it reads those too. Anthropic calls this progressive disclosure, and it means an agent can carry dozens of skills without filling its context window.
In December 2025 Anthropic published the format as an open standard at agentskills.io, so a skill written for Claude can work in other agent products. Same pattern as MCP: build the thing, then give it away so it becomes the default.
I'll say the uncomfortable part flat. Skills make a lot of software less necessary. If a workflow can be described in a text file and an agent can execute it, the case for buying a tool whose main value is "we encoded this process" gets weak. That transition will be messy and slow (enterprise software does not disappear in a quarter), but I'd bet on the direction.
Write down one workflow before you build an agent
For most teams the bottleneck sits upstream of the agent: nobody has ever written down how the work actually gets done. Pick one workflow your best person does well. Have them explain it step by step, including the judgment calls. That document is most of a skill. The rest is examples, a template and a way to check the output.
Agents Are Moving Off the Laptop and Into the Background
Most people still meet Claude in a chat window. You ask, it answers, you copy the result somewhere. That's changing.
The Claude Agent SDK lets teams run the same agent loop that powers Claude Code inside their own applications, with subagents, hooks, permissions, MCP and skills built in. And this spring Anthropic opened Claude Managed Agents in beta: a hosted harness where you define the agent once (model, prompt, tools, MCP servers, skills) and Anthropic runs the sessions, including long-running tasks and scheduled runs on a cron.
Put those together and the shape of 2027 looks like this. The weekly pipeline report builds itself Sunday night. The intake queue gets triaged before anyone logs in. The month-end reconciliation runs and flags the three items a person needs to look at.
Background work is where the real time savings are. It's also where things go wrong quietly. An agent that runs at 2 a.m. and makes a bad call doesn't have a person watching it.
Decide where each agent runs, and who owns it
For every agent that runs without a person in the loop, write down three things: where it runs, what triggers it, and whose name is on it when it breaks. If you can't answer the third one, it's not ready to run unattended.
The Digital Workforce Is a Management Problem Before It Is a Tech Problem
"Digital workforce" sounds like marketing (and a lot of the time, it is). But the idea underneath it is real, and leaders are already planning around it. In Microsoft's 2025 Work Trend Index, 82% of leaders said they're confident they'll use digital labor to expand workforce capacity in the next 12 to 18 months. The same report describes a new role it calls the "agent boss": someone who builds, delegates to and manages agents.
I like that framing more than I expected to. The hard part of an agent workforce looks a lot like the hard part of a human one. Someone has to define the job. Someone has to check the work. Someone has to decide when an agent gets more responsibility and when it gets pulled back.
The model handles the doing. Your people handle the deciding. Getting teams comfortable with that split, and willing to change how they work around it, is most of the project. That's the part that takes real effort, and it's the part most agent roadmaps skip.
Give every agent a manager
Treat a new agent like a new hire on probation. It gets a named owner, a narrow scope, a review of its work every week for the first month, and a clear rule for which actions need a person's sign-off. Expand its scope when the reviews earn it.
Workflow Logic Gets Cheaper While Context Gets More Valuable
This is the shift that matters most to me. When every company has access to the same models, the same protocols and the same runtimes, none of those are an advantage. They're table stakes.
What's left is context. Your data, your systems and, most of all, your written-down judgment about how the work should be done. Two companies running identical agents on identical models will get very different results if one has clean connectors into its data and well-written skills, and the other doesn't.
That's why the two middle layers in the stack below matter most. The model is shared by everyone. The runtime and governance layers are becoming standard products. Skills and connectors are where your company shows up.
It also changes what skills inside a company are worth. Knowing how to click through a system gets less valuable every quarter. Knowing what to ask for, and what good output looks like, gets more valuable. Judgment appreciates. Execution depreciates.
Invest in the context layer your agents read
If you have a budget line for agents in 2027, put a real share of it into the layer underneath them: connectors into the systems that matter, data clean enough to trust, and a library of skills your teams actually maintain. A better agent on top of messy context still gives you messy output.
Five Bets for 2027, and What I'd Do About Each
Here's how I'd sum up the trends. They're bets. They're also the ones I'd plan around.
| Trend | What it looks like in 2027 | What to do now |
|---|---|---|
| MCP everywhere | Every major system ships an MCP server, and agents connect to them by default | Build connectors once, scope them by role, reuse them across agents |
| Skills as company know-how | Teams keep libraries of versioned skills the way they keep templates today | Write down your best person's workflow and turn it into a first skill |
| Background agents | Agents run on schedules and triggers as well as in chat | Name an owner and an approval rule for every unattended agent |
| Agent managers | "Manage the agents" shows up in real job descriptions | Assign agent owners and give them a weekly review habit |
| Context as the moat | Results depend more on your data and skills than on the model | Fund the connector and skill layer before adding more agents |
Trust Is the Part Nobody Has Solved
I want to be honest about what's still open. We do not yet have a great answer for how humans and agents should share work day to day. The interface is still mostly a chat box. The security model for agent access is still being worked out, company by company. And trust calibration (when do you let an agent act without checking?) is mostly gut feel right now.
I wouldn't wait for those answers. I'd build the guardrails in from the start and skip the version where you bolt them on after an incident. Approval gates for anything sensitive or irreversible. Evals that check outputs against a known-good standard. Telemetry that shows what each agent did and why.
Build approval gates in from day one
Before an agent goes live, list every action it can take and mark which ones need a person to approve. Sending an email to a customer, changing a record, moving money: those get a gate. Reading, drafting and summarizing usually don't. Loosen the gates as the evals and the track record earn it, not before.
The Workforce Gets Built One Workflow at a Time
I'm bullish on where this goes. The standards are in place, the runtimes are real, and the cost of trying an agent on a real workflow keeps dropping. 2027 is the year a lot of companies move from "we have Claude" to "Claude does part of the job every day."
What I don't know is how fast the management side catches up. How many agents can one person really oversee? What does a performance review for an agent look like? Which work should never be handed off? I don't think anyone has clean answers yet, including us.
What I'm sure of is that the companies that get there first will start small: one workflow, written down well, connected to the right data, with a person who owns it. That's how we approach PressW's Custom Claude Agents work: scoped skills, agents and MCP connectors built around how your team already works, with evals, telemetry and human approval gates included. If you're trying to figure out which workflow to start with, I'm happy to think it through with you. Book a Discovery Call.
FAQ
The main trends are MCP becoming the default way agents connect to business systems, Agent Skills becoming the way companies package their know-how, agents moving into scheduled background work, and new roles for people who manage agents. Underneath all of them, company context (data, connectors and skills) becomes the main source of advantage.
MCP is an open protocol, introduced by Anthropic in November 2024, that gives AI agents a standard way to connect to tools and data sources. It is now governed by the Agentic AI Foundation under the Linux Foundation and is supported by most major AI products. For custom Claude agents, MCP connectors are how the agent reads and writes to your CRM, data warehouse and document stores.
Agent Skills are folders that contain a SKILL.md file with instructions, plus optional scripts and reference files, that teach an agent how to do a specific task. Claude loads only a skill's name and description until a task needs it, so an agent can carry many skills at once. The format is an open standard, so skills can work across agent products.
A digital workforce means AI agents that handle defined pieces of work, like triage, reporting or first drafts, alongside your team. For a mid-size company it usually starts with a few agents, each with a narrow job and a named owner. The management practices (scope, review, approval rules) matter as much as the technology.
Vendor agents work well for tasks that look the same at every company. Custom Claude agents make sense where your workflow, data or judgment is specific to you, because that's where a generic agent falls short. Many companies will run both.
Scope every MCP connector by role, put approval gates on sensitive or irreversible actions, run evals on agent outputs, and log what each agent does. Give every agent a named owner. These controls are easier to build in at the start than to add after something goes wrong.
