Skills + Search2o
Build each repeatable task on your company's systems as an agent in Search2o, and install one skill in the agent harness. The skill makes two calls on your agent server: search finds the agent for a request, and execAgent runs it. The agent harness keeps the agentic loop. The work on your systems runs on the agent server.
The usual alternative is a skill per workflow, with MCP servers or scripts that reach company systems from each laptop. There, Claude reads the instructions and writes and runs the steps itself, on every run, in every person's session.
The table lists what changes when those workflows become Search2o agents. What Search2o adds describes each change, with its sources.
At a glance
| What changes | A skill per workflow, with MCP servers or scripts | One skill, agents in Search2o |
|---|---|---|
| Where a workflow lives | In instructions Claude interprets on each run | In a fixed, validated, versioned agent that runs on your server |
| What enters Claude's context | The instructions, every step Claude writes, every intermediate result | The skill's instructions, the names and titles of matching agents, each agent's output |
| Model tokens per run | Finding the skill, loading it, and a model call for each step Claude writes | One skill and two calls by Claude, plus the agent's llm steps; search and other commands cost none |
| What the model work costs | Claude usage for every step | Claude usage for the two calls; each llm step at the price of its model, or no per-token charge on a model you host |
| When the assistant's model changes | Each skill is due a retest | One skill is due a retest |
| Models a step can use | Claude models | Any vendor's, or one you host with Ollama, vLLM or SGLang, chosen per step in a profile |
| What a laptop holds | A route and a credential for each system, or for the MCP server in front of it | One revocable token for the agent server |
| What can act on your systems | What the model decides at run time, within the person's access and permission rules | Published agents, inside their profiles, allowlist, time and cost limits |
| Record of a run | Session telemetry, if you collect it | Agent, version, person, duration, result and cost, in built-in reports |
| Adding a workflow | A skill has to reach each person's machine or account | An agent is published |
What stays with Claude
The agent harness keeps the agentic loop. It splits a request into parts, sends each through search and execAgent, decides the order and the conditions from the results, and composes one answer.
A support engineer types: Has order 48812 shipped? If not, open a ticket with the warehouse. Claude runs the order-lookup agent, reads "not shipped", then finds and runs the warehouse-ticket agent in the same conversation. Search2o did each step inside the company. Claude decided the order and the condition.
Claude also keeps everything it does without Search2o: files, code, and requests nobody has built an agent for. When nothing fits, search returns no match instead of guessing. In published tests it refused 100% of unrelated questions and 82 to 86% of in-domain questions that no agent covered. Claude then answers as it normally would.