Clare Liguori

Clare Liguori is a senior principal engineer at Amazon Web Services working on Kiro and the Strands Agents SDK. Her work spans automated software delivery and agentic developer tools, with a recurring engineering question: what must a system be able to check and correct before people can trust it to work with less supervision?

Developer tooling and safe, hands-off deployment

Liguori studied computer science at the University of Texas at Austin and worked on WiFi products at Motorola Solutions before joining Amazon in 2014. At AWS, she worked on the AWS Code Suite, Amazon ECS, and AWS Proton before moving into agentic AI. Her career and project history connects developer tooling with the infrastructure needed to deploy and operate software.

Her open-source contributions include AWS service modules in the Cloud Development Kit and Amazon ECS integration in Spinnaker, where she previously co-led the AWS special-interest group. In Backstage, she contributed a library that standardizes how plugins authenticate to AWS APIs and maintained the AWS Proton plugin. She also published GitHub Actions for continuous deployment to ECS. These contributions address practical gaps between an application and its delivery environment: provisioning services, authenticating developer tools, and moving changes into production.

Her writing on safe, hands-off deployments explains how automation can limit the damage from a faulty release. Small rollout stages restrict initial exposure; monitoring and automatic rollback respond to failures; waiting periods give delayed problems time to appear before a change spreads. In her account of automated release management, pipelines also preserve the relationships between immutable artifacts, source commits, and deployment history. Hands-off operation depends on checks, visibility, and a way to recover.

Strands: model-directed agents with targeted checks

In early 2023, Liguori and her Amazon Q Developer team began building agents for software development and operational troubleshooting. Early models needed extensive prompting, response parsing, and orchestration to use tools reliably. As models gained native tool-use capabilities, some of that scaffolding constrained their behavior. The team developed Strands around a simpler combination of a model, tools, and a prompt, allowing the model to choose its next action. Liguori introduced the open-source SDK in May 2025, when several AWS teams were already using it in production.

• Steering hooks: That preference for model-directed execution still leaves room for explicit safeguards. In her work on Strands steering hooks, handlers inspect behavior before tool calls and after model responses. They can check parameters and prior actions, block an invalid operation, and return feedback that lets the model correct itself. In a library-book renewal experiment, she reported a 100% pass rate across 600 evaluation runs, compared with 82.5% for the tested prompt-instruction approach and 80.8% for graph-based workflows. The result supports targeted checks in that experiment; it does not establish a universal reliability guarantee.
• Retrieval as a tool: Liguori also advocates giving agents retrieval as a tool, so they can formulate queries when they need information. Loading material into the context window in advance can introduce irrelevant content that distracts the model. She credits colleague Anoop Deoras with coining the “RAT” shorthand for this approach.

Shared infrastructure for agent execution

As a Kiro engineering lead, she helped consolidate separate IDE, CLI, and web implementations into a unified agent harness. The co-authored account describes a standalone server process that separates agent execution from each client’s interface, with Agent Client Protocol providing communication between them. Shared tools, session handling, and configuration reduce duplicated implementation and inconsistent behavior. The design also supports the ambition of continuing work between local and cloud environments.

In April 2026, Liguori joined the Model Context Protocol core maintainer group. The announcement described her contributions to discussions about unsolicited tasks, agent execution, and triggers and events, extending her developer-tooling work into shared infrastructure for agent interactions.

What Amazon’s agent-driven development experiments show

Her 2026 account of frontier development examines what changes when agents perform most implementation work. She describes developers who give agents enough direction to run for hours, validate their own changes, and work through tasks in parallel. Her argument is that sustained gains require teams to change their engineering practices, rather than simply add a coding assistant to an existing workflow.

The Amazon experiments she discusses show both the potential and the limits of that claim. She reports that the Bedrock Mantle team built a new inference data plane with six engineers in 76 days, against an earlier estimate of 30 people over 18 months. She also emphasizes that this was an unusually experienced team, including two distinguished engineers. A Prime Video experiment reduced a delivery estimate from 90 weeks to 24 after a ten-day sprint, but the engineers had few distractions and a senior engineer had spent three weeks preparing detailed tasks. The revised estimate was not a completed delivery, and the sprint’s conditions differed from ordinary day-to-day work.

A broader Amazon Stores pilot followed 50 teams working on existing systems and measured deployment velocity to production. Liguori describes stronger gains among teams that deliberately changed how they worked. These internal observations support her emphasis on working practices, while leaving open how readily the results transfer to other teams and organizations.

Preparing teams and codebases for autonomous work

• Knowledge and codebase readiness: She makes that preparation concrete. Teams need to write down knowledge previously passed through onboarding, conversations, and reviews, then update agent instructions when mistakes reveal missing context. They also need to remove instructions that newer models no longer require, so useful guidance does not become context bloat. Codebase readiness includes navigable structure, informative tool errors, and tools or MCP servers that let agents complete their tasks. She describes teams adopting TypeScript or Rust to give agents more useful compiler feedback, without treating a language change as a requirement.
• Autonomous validation is central to her approach. An agent needs clear requirements and a quality bar it can check through compilation, linting, and tests. For ambiguous features, she favors settling intent in a specification before generating changes across a codebase; a document is easier to revise when the underlying requirements are still wrong. Fast local tests and mock services with deterministic responses let agents detect mistakes and retry without repeatedly waiting on remote infrastructure. Those feedback loops make longer unattended work possible.

Managing the costs and workload

Liguori is equally explicit about the costs. Teams often slow down initially while improving their codebases and learning new habits. Running several agents increases context switching, and reviewing generated code can impose particular strain on early-career engineers who have less review experience. She warns about burnout and about rolling out new practices across an organization before smaller experiments have established what works. As implementation becomes faster, product decisions and launch reviews can become the longest parts of delivery; she argues for faster decisions, especially when they are easy to reverse.

Her earlier Clare’s Plate document addresses a related problem in human engineering work: making responsibilities visible enough to discuss priorities. It groups work into Current, Back Burner, and Done, including mentoring and design reviews that sprint boards can miss. Sharing it with her manager creates room to discuss blocked work, focus, and career development. That attention to preparation, feedback, and workload remains an important part of her account of useful automation.