DEV Community: HIVE

What does an agent actually need from your app to do one useful thing

HIVE — Wed, 14 Oct 2026 12:28:26 +0000

We started with an assumption about what agents need from an app. Then we started asking developers. The answers keep surprising us. So here is the question we are testing now. What does an agent actually need from your app to do one useful thing. Tell us the workflow. Tell us where it breaks.

What we're learning building HIVE

HIVE — Wed, 14 Oct 2026 03:06:06 +0000

The hardest objection we have heard this week came from three separate threads. Structured access is ideal. But you do not own the other side.

That objection is now the specification. The clean path is the vendor handing the agent tidy actions. The real path is the software never designed for agents at all. Dashboards, forms and workflows built for human eyes only.

This week practitioners wrote the requirements for us. Testimonials need a human before they go live. Escalation happens after two contacts. Refunds under fifty dollars are fine when the order is paid and never refunded before. Timeline rules nobody wrote down anywhere.

None of that is in any API contract. All of it is what an agent has to get right. The structure agents need is not technical. It is operational.

And the market keeps proving the demand. This month alone, more than a dozen vendors shipped their own agent doors. Blockworks, notolens, Genea, Tinuiti, Prembly, Silverchair and Microsoft's Business Central team. The vendors who own their software can build this themselves. The thousands of SaaS products whose vendors never will are the layer we are building.

What is the one unwritten rule in your workflow that no API documents and every agent would get wrong.

AIAgents #AgenticAI #BuildInPublic #DevCommunity

One integration is a project. Ten is a team. Fifty is a department that only maintains connections.

HIVE — Tue, 13 Oct 2026 12:30:46 +0000

One integration is a project. Ten is a team. Fifty is a department that only maintains connections. Every app brings its own quirks, its own auth, its own failure modes. The work multiplies even when the apps do not change. At what number of integrations did maintenance become someone's full time job.

Why build another integration per app?

HIVE — Tue, 13 Oct 2026 03:13:27 +0000

Every SaaS company is now building its own agent door. Blockworks shipped a hosted MCP server for its datasets. Notolens built one for brand protection. Genea for access control. An insurer launched an entire directory so every carrier can list its own MCP server. Silverchair ran a six-month pilot of one shared endpoint across independent publishers instead of one per publisher.

Stacklok's 2026 report says 45 percent of software organizations already run MCP servers. That is the right instinct. Agents need structured doors, not scraped pages.

But count the cost. Every door is a separate integration to build, secure, version, and maintain. Every API change is an agent breakage waiting to happen. Every new tool in the stack starts the work again from zero. The industry is rebuilding the same connector dozens of times and calling it agent strategy.

One integration per app scales linearly. One infrastructure layer scales once.

If every tool you buy ships its own agent interface next year, who maintains the fifty doors in your stack.

AIAgents #AgenticAI #MCP #DevCommunity

What a good CLI walkthrough should make visible

HIVE — Mon, 12 Oct 2026 12:29:57 +0000

A good CLI walkthrough shows everything. The exact command. The inputs it took. The action it performed. The output it returned. The errors it can throw. And how you verify it worked. No cropped terminals. No trust me it worked. What is the first thing you check when a walkthrough skips the verification step.

The terminal is the most honest place to test an agent

HIVE — Mon, 12 Oct 2026 03:04:29 +0000

The most honest place to test an agent is the command line. There is no dashboard to hide behind.

Take a routine request. Find every customer whose trial ends this week and draft the renewal email. Three steps. Read the records. Filter by date. Write the draft.

Today this runs through a terminal with whatever permissions the environment happened to grant. Reading one record and reading every record are the same door. The write step that sends words to a customer has no approval pause. The permission decision is invisible. After the fact, the log says the command ran. It says nothing about whether it should have.

A CLI-driven task through a contract layer looks different. Structured reads with field-level scoping, so one record and every record are different doors. The filter runs against typed records, not scraped output. The write step waits for a named human before anything leaves.

The walkthrough is short on purpose. The point is not the commands. It is that every step in the transcript carries its permission decision, visible in the log, auditable after the fact.

When an agent runs a task from your terminal, where is the permission decision recorded. And who approved it.

AIAgents #AgenticAI #CLI #DevTools #DevCommunity

A guardrail you cannot test is a hope

HIVE — Sun, 11 Oct 2026 12:33:48 +0000

A guardrail you cannot test is a hope. Try the unauthorized action and watch it stop. Try the missing field and watch it ask. Try the stale data and watch it refresh. Try the ambiguous request and watch it clarify. Try the step that needs approval and watch it wait. Which test would your current setup fail.

Agents with guardrails

HIVE — Sun, 11 Oct 2026 03:08:55 +0000

Two enterprise security vendors shipped the same message within days of each other. Agents need guardrails. Almost nobody has them yet.

RSA launched Agent ID at the end of September. Every agent call passes through a policy gateway. Authorization goes down to the tool and the argument. High risk actions like wire transfers, restricted data and payments wait for a named human to approve through a separate phishing resistant channel before anything moves. Each approval creates a record tied to the person who authorized it.

Days later SailPoint expanded its Agentic Fabric to find shadow agents. Agents nobody registered. It applies runtime authorization and can disable one on the spot while legitimate agents keep working.

Their own research explains the urgency. 79 percent of enterprises already run AI agents in production. About 2 percent have identity security tools designed to govern them.

The demand side is settled. The identity vendors are building the contract layer for agents. The open question is the software with no identity team behind it.

That is the layer we are building with HIVE. Permissions designed into the access layer from day one. Not discovered after the fact.

When an agent asks to move money in your product, who says yes. And how would you prove it.

AIAgents #AgenticAI #IdentitySecurity #DevCommunity

The hard part is not connecting systems. It is predictable behavior.

HIVE — Sat, 10 Oct 2026 12:30:02 +0000

Connecting systems is the easy part. The hard part is predictable behavior. The same request should do the same thing every time. Edge cases should fail the same way twice. Permissions should hold under pressure, not just in the demo. Error handling should tell you what happened, not just that something did. Which of these breaks first when you connect a new system.

The hardest problem we're solving

HIVE — Sat, 10 Oct 2026 03:05:59 +0000

The most repeated pushback we have heard this week did not come from one thread. It came from three. Structured access is ideal. But you do not own the other side.

It is a fair objection. And it is the exact problem we are working on. Not the happy path where a vendor hands the agent clean actions. The software that was never designed for agents at all. Dashboards, forms and workflows built for human eyes only.

This week practitioners gave us the shape of it in their own words. Timelines with rules nobody wrote down. Testimonials that need a human before they go live. Escalation after two contacts, not one, not three. A refund under $50 that is fine when the order is paid and has not been refunded before.

None of that is in any API contract. All of it is what an agent has to get right.

So here is the question we are stuck on. When the software will not give the agent structure, where should the structure come from. What would you build first.

AIAgents #AgenticAI #BuildInPublic #DevCommunity

A tool description should answer five questions

HIVE — Fri, 09 Oct 2026 14:06:24 +0000

A tool description is a contract between the builder and the agent. A good one answers five questions.

When to use it. Not a feature list. The situations where this tool is the right call, and where it is not. Without this, the agent reaches for it everywhere.

What inputs it requires. Every field, what it means, which ones are mandatory. Vague inputs are where silent failures begin.

What result to expect. What success looks like when it comes back. An agent that cannot recognize success cannot check its own work.

What errors it can throw. Every failure mode, stated plainly. An agent meeting an undocumented error will misread it.

What permission rules apply. Who may call it. What needs approval. What is off limits. This is the line between useful and dangerous.

Five questions. Most tool descriptions answer one. Which one does your team skip most often.

The agent stops improvising: roles and business logic

HIVE — Fri, 09 Oct 2026 03:04:57 +0000

So far in this series: the translation layer turns your software into structured actions, and permissions plus approvals decide what each agent may touch.

Part three ties it to how your team actually works: roles and business logic.

Roles are how you stop configuring agents one by one. A support role bundles search-records, draft-response, and escalate-ticket. A billing role bundles read-invoice and issue-refund-under-threshold. Define the role once, assign it to any agent — human, internal copilot, or third-party model — and the permission boundaries travel with it. Change the role, and every agent wearing it updates at once.

Business logic is the part most agent setups skip. Your company doesn't run on "create ticket." It runs on rules like: refund under $50 approves itself, over $50 needs a manager; enterprise tickets route to the senior queue; nothing ships until the order is paid. HIVE lets those rules live where the actions live. The agent doesn't have to learn your playbook from a prompt — the playbook is part of what the software exposes.

That's the full picture of how it works. Actions are the vocabulary. Permissions are the boundaries. Roles are the job descriptions. Business logic is the way the company already runs.

Put together, the agent stops improvising and starts operating the way you intended.

Which of your team's unwritten rules would you encode first?

AIAgents #AgenticAI #BuildInPublic #DevCommunity #AI #SoftwareEngineering