StackLift
All terms

Glossary

AI agents (in software delivery)

Specialized AI workers (scope, code, QA, DevOps...) that draft work for human review.

AI agents in software delivery are specialized, role-scoped assistants - for scoping, architecture, code, QA, DevOps, documentation, review, and PM - that accelerate a delivery phase by drafting work which senior engineers then review and approve.

Used well, agents need a hierarchy and explicit permissions: a master agent coordinates, leader agents own workstreams, and specialist agents execute narrow tasks, escalating anything ambiguous.

Agents never push to production unattended in a governed setup; they open pull requests and surface drafts for human decisions.

Start in 60 seconds

Put this to work on your project.

Get an AI-assisted, architect-reviewed scope, cost range, and timeline. Free, no login.

The complete guide

AI agents (in software delivery): the complete guide

Updated November 18, 2026 ยท Published by , Founder, StackLift AI. Written with AI assistance; how we make content.

Short answer

AI agents are software components built on a language model that can plan steps, use tools such as databases and APIs, and act toward a goal. In software delivery they are specialized, role-scoped assistants for scoping, code, QA, DevOps and documentation that draft work for senior engineers to review and approve.

What are AI agents?

An AI agent is a program that uses a language model to decide what to do next, then does it. Unlike a chatbot that only answers, an agent can call tools: search a knowledge base, read a record, run a query, write a file or draft a message. It looks at the result, decides whether the goal is met and continues if not.

Three parts define an agent: a model that reasons, a set of tools it is allowed to use, and a loop that lets it act and check. Instructions and limits wrap around them. The loop is what makes it an agent; the limits are what make it safe to use.

The wider approach is called agentic AI. The agent is the component, and agentic AI is the way a whole system is designed around such components.

What are AI agents in software delivery?

In software delivery, agents are role-scoped assistants that speed up one phase of building software. One drafts scope and user stories, another proposes architecture, others write code, tests, infrastructure files, documentation and review notes. Each has a narrow job and defined permissions.

Their output is a draft. Senior engineers review and approve it, and people own the decisions. This is the idea behind an AI software factory: agents accelerate the work, and a governed pipeline keeps quality, security and control in human hands.

Typical delivery agent roles
RoleWhat it draftsWho reviews
ScopingRequirements, user stories, estimatesArchitect or product lead
ArchitectureSystem design options and trade-offsSenior engineer
CodeImplementation on a feature branchEngineer through code review
QATest cases, test code, defect reportsQA lead
DevOpsPipeline and infrastructure definitionsDevOps engineer
DocumentationGuides, release notes, API docsOwner of the module
ReviewFirst-pass code review commentsHuman reviewer
Project managementStatus summaries and risk flagsDelivery manager

How do AI agents work, step by step?

  1. The agent receives a task with context, such as a ticket, a document or a question.
  2. It reasons about the goal and decides on a first action, sometimes writing a short plan.
  3. It calls a tool through tool calling, such as a search, an API request or a file edit.
  4. It reads the tool result and updates its working context.
  5. It checks whether the goal is met, whether a rule says to stop or whether it should ask a person.
  6. It repeats until done or until a limit on steps, time or cost is reached.
  7. It returns the result with a log of what it did, for review.

Between steps, a control layer enforces limits and approvals. See agent orchestration for how that layer is built, and agent memory for how agents keep information across steps.

What types of AI agents are there?

Agents can be grouped by how much they decide and how they are arranged. These categories help you pick a design.

Types of AI agents
TypeHow it behavesGood for
Tool-using assistantAnswers and calls a few tools on requestLookups, drafting, simple actions
Workflow agentFollows a defined path and uses a model at certain stepsPredictable business processes
Autonomous agentPlans its own steps toward a goal within limitsOpen-ended research or troubleshooting
Multi-agent teamSeveral agents with different roles and a leadLarge tasks needing different skills; see multi-agent system
Coding agentReads a codebase, edits files, runs testsSoftware tasks on branches with review

More autonomy is not better by default. The more freedom an agent has, the more you need testing, limits and review.

How should AI agents be organized and controlled?

Used well, agents need a hierarchy and explicit permissions. A common pattern is a coordinating agent that breaks work into parts, leader agents that own workstreams and specialist agents that execute narrow tasks. Anything ambiguous or risky escalates upward, and finally to a person.

  • Give each agent only the tools its job requires.
  • Keep credentials separate per agent and never expose production secrets.
  • Require pull requests or draft states instead of direct changes.
  • Add approval steps before money, customer messages or irreversible actions.
  • Log every action so a reviewer can trace what happened.
  • Set limits on steps, time and spend.

In a governed setup, agents never push to production unattended. They open pull requests and surface drafts for human decisions. The post on the agent delivery hierarchy describes this structure in more detail.

What are examples of AI agents in business?

  • Support: read a ticket, find the policy, draft a reply and escalate refunds above a limit.
  • Finance: extract invoice data, match to orders and flag mismatches for approval.
  • Sales: research a lead, update the CRM and draft an outreach email for review.
  • Legal and compliance: scan contracts against a clause list and flag items for a lawyer.
  • IT: triage requests and run approved fixes such as access resets.
  • Operations: monitor inventory or shipments and propose reorders or reroutes.

Browse AI agents by industry for designs in healthcare, logistics, real estate and more, and the cost pages for project ranges.

AI agents vs. chatbots, copilots and automation

AI agents compared with related tools
ToolWhat it doesWho actsBest for
ChatbotAnswers questions in conversationThe userFAQs and information
AI copilotSuggests inside a productThe user accepts or editsAssisting people at work
AI agentPlans and takes actions through toolsThe agent, within limitsMulti-step tasks
RPA botReplays recorded stepsThe bot, on fixed pathsStable, repetitive screens
Workflow automationMoves data on rules and triggersThe rule engineWell-defined processes

Compare in depth: AI agent vs. chatbot, AI agent vs. RPA and single-agent vs. multi-agent.

What are the risks of AI agents?

Risks and controls
RiskControl
Wrong or unsupported outputGround in sources; test for hallucination; human review
Actions beyond intentLeast-privilege tools; approvals; reversible steps
Injected instructionsTreat content as untrusted; see prompt injection
Runaway costLimits on steps and spend; monitoring
Data exposureScoped access, redaction and logging
Quality driftRegular agent evaluation and monitoring

The thread through all of these is that people stay accountable. See human-in-the-loop and AI guardrails.

How do you start with AI agents?

  1. Choose one task that is frequent, multi-step and checkable.
  2. Write what good looks like, what the agent may touch and where a person must approve.
  3. Build with few tools and approvals switched on.
  4. Test on real cases and awkward ones before anyone relies on it.
  5. Run it beside people first, comparing its drafts with their work.
  6. Launch with logging and a way to switch it off.
  7. Review results regularly and widen scope only when evidence supports it.

The AI agent evaluation checklist and agentic workflow design playbook cover these steps in detail.

How do you measure whether an AI agent is working?

Measure the agent the way you would measure a new team member: by the quality of its work, how often it needs help and what it costs. Qualitative signals are enough to start, as long as you check them against real cases on a schedule.

  • Task success on a fixed test set of realistic and awkward cases.
  • How often a person must correct or redo the agent's output.
  • Whether it used the right tools and stayed within its permissions.
  • How often it escalated, and whether those escalations were right.
  • Time and spend per completed task, tracked as volume changes.
  • User feedback from the people who rely on the output.

Keep traces for every run. When something goes wrong, a trace shows which step failed, and that is usually enough to fix it.

When are AI agents not the right choice?

Agents add variability and review work, so use them only where that is repaid. If a process is fixed and rule-based, workflow automation is cheaper and easier to verify. If a single answer is needed, a chatbot or a plain model call is enough.

  • Mistakes are costly and cannot be reviewed or reversed.
  • There are no test cases or data to judge quality.
  • The systems the agent would need are unreliable or have no usable interface.
  • The task is rare, so building and testing an agent costs more than doing the work.

Fix the process or the data first, then reconsider. See workflow automation for the simpler route.

How do you choose a partner to build AI agents?

Ask how they scope, test and control agents, not only what models they use. A good partner can show you how limits, approvals and logs are designed before any code is written.

  • Do they run structured discovery and review scope before pricing?
  • How do they test agent behavior, and will you see the results?
  • What permissions will the agent have, and how are secrets handled?
  • Where do people approve actions, and can you change those points?
  • Will you own the code, the prompts and the test set?
  • How will they hand over monitoring and maintenance?

What tools and platforms are used to build agents?

Models come from OpenAI, Anthropic Claude and Google Gemini, among others. Frameworks include LangGraph, the OpenAI Agents SDK, CrewAI and the Vercel AI SDK. The Model Context Protocol is an open standard for connecting agents to tools and data. Cloud services such as AWS Bedrock, Azure AI Foundry and Google Vertex AI host and manage agents.

Supporting parts include a database or vector store, an observability stack for traces and a CI/CD pipeline. Pick for control features like permissions, tracing and approvals, not only for model quality.

What do AI agents mean for cost and timeline?

Cost has two parts: building the agent and running it. Building depends on tools, integrations, approval flows and testing. Running depends on steps per task, context size and model choice, since each step is a model call.

For current model prices see the LLM pricing page, and for ranges see AI agent costs and AI agent cost by type. The free estimate scopes your own build. Most first versions take 6 to 16 weeks depending on scope.

How does StackLift use AI agents?

StackLift is an AI-assisted software delivery partner: an architect reviews scope, then senior engineers and AI agents ship it. Agents work on feature branches and never push to the default branch or touch production secrets. Every estimate is reviewed by a senior architect before it becomes a commitment.

Clients see live project state, approvals, invoices and releases in the client portal. If you want agents built for your business, see AI agent development.

When you want agents running in your own business, not only used in our delivery, our forward deployed engineers build and launch them in your systems, starting from a 3-day scoping plan.

Questions people ask

What is the difference between an AI agent and a chatbot?

A chatbot answers questions in a conversation. An agent can also plan steps and take actions through tools, such as updating a record or opening a pull request, and checks its own results as it goes.

Do AI agents replace developers?

In a governed setup, no. Agents draft code, tests and documents, and senior engineers review and approve them. People own the decisions and the quality bar.

Are AI agents safe to use with company data?

They can be, with limits. Scope access to what the task needs, keep secrets away from agents, log actions and require approval for risky steps. Safety comes from design and oversight.

What can an AI agent actually do?

It can search and read data, call APIs, write and edit files, send drafts and run checks, depending on the tools you give it. It cannot do anything its tools and permissions do not allow.

How is an AI agent different from automation?

Automation follows fixed rules. An agent interprets inputs and picks its next step, so it copes with variation. That flexibility needs more testing and oversight.

How many agents do I need?

Often one. Add more only when tasks need different tools, permissions or skills. Extra agents add cost, delay and more places for errors.

How do I test whether an agent works?

Build a set of realistic and awkward cases, run the agent on them several times and score the results and the steps. Repeat after every change. This is called agent evaluation.

What does it cost to build an AI agent?

It depends on tools, integrations, approvals and testing. Use the cost pages and the free estimate for ranges, and the LLM pricing page for model running costs.

Sources

Sources and further reading

The standards, rules, research and official documentation this page draws on.