Elvis Saravia Improves Agentic Loops Reliability with HITL and DialAgent

Elvis Saravia now routes every Claude and Codex agent session through DialAgent to add human oversight inside agentic loops. The change lets agents escalate decisions to people via short calls when automation handles pull requests or new features.

The setup turns a phone number into a native tool for each agent. Voice, SMS and iMessage become direct channels that the agent can trigger without extra code. When the loop reaches an ambiguous choice, the agent calls or texts the designated human instead of guessing or stalling.

Saravia keeps the calls brief, especially when he is away from a desk. The agent states the current decision point, waits for a verbal reply, and then continues the task. This pattern replaces long waits or silent failures that often appear in fully autonomous runs.

Users can paste a single instruction into the agent chat and ask the agent to place a test call. The flow confirms the MCP server connection and shows how escalation would work inside a live workflow before the agent runs unattended.

DialAgent supplies a five-dollar starting balance so developers can try the voice path without committing funds up front. The service connects to existing agent frameworks instead of requiring a new platform.

Agentic loops gain a safety valve

Most agentic loops still run into edge cases that code alone cannot resolve. An agent might misread a merge conflict, choose the wrong dependency version, or loop without reaching a stable state. Adding a phone-based escalation point gives the loop an exit that leads back to a human rather than an error log.

Saravia applies the method to two common tasks: processing pull requests and building new features. In both cases the agent follows its normal steps until it needs approval or clarification. The call happens only then, keeping the human load low while keeping the loop moving.

DialAgent turns communication into a tool

The MCP server registers a dedicated number for each agent. The agent can send voice, text or iMessage without leaving its normal tool-calling format. This removes the need to build separate notification systems or rely on email queues that sit unread.

Because the channel is native, the agent treats the call the same way it treats a file-write or search operation. The response from the human returns as a message the agent can parse and act on immediately. The loop therefore continues with updated context instead of restarting.

Human oversight stays lightweight

The calls stay short by design. Saravia reports that a typical escalation lasts under a minute. The agent states the choice it faces, receives a one-sentence answer, and returns to its task. This keeps the human time cost small while preventing the agent from making silent mistakes that later require hours of debugging.

The approach works well for developers who move between locations. When an agent reaches a decision point on a weekend or during travel, the phone route reaches the person without requiring them to open a laptop.

Testing the escalation path before production

Developers can paste a test instruction such as “Call me and ask whether to update the dependency.” The agent places the call, speaks the question, captures the reply, and reports the outcome back in the chat. The test confirms both the connection and the parsing step without touching real code.

This verification step lowers the risk that the escalation path fails at the moment it is most needed. It also gives new users a concrete example of how the human-in-the-loop handoff works inside an otherwise automated sequence.

Limitations remain visible

The method still depends on the human answering the call. If the person is in a meeting or has the phone muted, the loop pauses. Saravia has not published data on average wait times or missed escalations, so the actual reliability gain is not yet quantified beyond individual use.

In addition, the $5 credit covers only initial testing. Teams that scale the approach to many agents will need to budget for ongoing usage. No public pricing tiers beyond the starting balance have been released.

What to watch next

Teams that adopt similar phone-based escalation will likely share usage logs or success rates within the next three months. Watch for public posts from other developers who route agents through DialAgent and report loop completion percentages before and after the change.

DialAgent may release additional tool integrations or log exports that make the escalation data easier to analyze. Any such release would let practitioners measure whether the short-call pattern improves reliability across larger codebases.

Finally, competing services could offer comparable voice channels inside agent frameworks. If a second provider appears with lower latency or cheaper per-minute rates, the cost-benefit picture for human-in-the-loop agent loops would shift again.

The pattern Saravia describes shows one practical way to add human judgment without turning every agent run into a full review session. Developers can test the route today with the provided credit and decide whether the added step fits their own automation needs.