Haiku 5.5 Is 75% Cheaper: What to Do With Agents in Production

AWS added Claude Haiku 5.5 to Amazon Bedrock.

According to Anthropic, this is the fastest and most efficient model in the 5.5 family.

It was designed for subagents and high-volume tasks where cost matters, and for most tasks it costs about 75% less than Haiku 4.5.

When model calls become four times cheaper, budget stops being the main constraint for agents in production.

Two questions remain: what data the agent is authorized to read and who is responsible for the fix it proposes.

A subagent is a focused executor within a larger system.

One agent reads the logs, another queries the database system views, and a third checks the result against the procedure.

Each one makes many short calls—thousands of them—so the bill is driven by cheap, fast models.

The roughly 75% price reduction compared with Haiku 4.5 changes the math: an architecture once considered too expensive for continuous operation can be reconsidered. Bedrock adds conditions to the model that the security team can verify.

Data stays within AWS infrastructure in the selected region, while access is controlled through IAM using the same roles the team already uses.

If the company is not ready to send internal data to an external API, the pilot cannot start without these conditions.

Cornerstone OnDemand serves 140 million users in 186 countries.

Previously, the DataOps team spent up to 45 minutes on each database incident, manually querying system views and comparing the results.

The Orion AI multi-agent system now runs on Amazon Bedrock and Strands Agents, AWS's open framework for agent orchestration.

Specialized agents investigate the incident in parallel.

The company reports a 78% reduction in diagnostic time.

With a 45-minute upper limit, that means about ten minutes instead of three quarters of an hour.

This is our calculation based on the stated figures; there is no independent measurement, so it should be applied with caution.

The mechanics are simple: the orchestrator splits the task, subagents perform focused checks, and the results are combined into a single conclusion.

A cheap, fast model acting as a subagent makes this architecture economically viable.

AWS DevOps Agent analyzes incidents around the clock using metrics, logs, and application topology.

It produces a root cause analysis (RCA) and recommended actions.

The on-call engineer receives a ready diagnosis in the middle of the night, but applies the fix manually.

A new AWS article describes the next step: automating remediation after the agent's investigation. This is where responsibility enters the picture.

The diagnosis changes nothing, while the fix changes production.

A workable design requires three things: a limited set of permitted actions; an audit log showing what the agent did and why; and a checkpoint where a person approves a risky step.

Without them, automated remediation becomes a source of new incidents.

At KT.Team, we design these safeguards during integration planning, before the first incident.

RAG (retrieval augmented generation) is a setup where the model answers using company documents retrieved from SharePoint, Google Drive, and Confluence. In these systems, each document has its own permissions list.

If an assistant answers using materials an employee is not authorized to view, that is a data leak, even when the answer sounds harmless. Amazon Quick with Amazon Bedrock Knowledge Bases checks permissions at request time: it compares the source ACLs and does not copy them into the index in advance. The difference is significant.

Copied permissions become outdated on the very day an employee moves to another department.

A check at request time sees the current state.

For CIS companies, the same principle applies to 1C,

Bitrix and corporate knowledge bases: the access layer determines whether RAG can be run on the data at all, so it must be built before choosing a model.

AWS separately describes how to close the gap between discussions about AI and building solutions.

In its assessment, the main adoption barrier is a lack of practical experience among professionals whose work is not related to coding but increasingly depends on AI.

They need tools, structured support, and room to make mistakes.

This matches our experience: effective agents emerge when the process owner helps build the solution together with engineers.