AI coding state of play, Agile meets Architecture, March 10 2026

Overview of the current state of AI coding tools, and questions technical leaders should be asking themselves right now.
- Context engineering
- Increased autonomy, reduced supervision - ups and downs
- Harnesses

Transcript

1. © 2026 Thoughtworks What AI augmentation means for technical leaders March 2026 1
2. © 2026 Thoughtworks AI relevance for technical leaders Use AI for your own work Help teams amplify the good, not the bad and the ugly Make it as safe as possible to use AI Use AI to improve and modernise the systems landscape
3. © 2026 Thoughtworks Context engineering: Beyond rules files and MCP servers
4. © 2026 Thoughtworks 4 - The backend code is Python, it's in ./app - Build system: Poetry - Remember to activate the virtual environment before any Python or Poetry command Simplest form: “Rules” files AGENTS.md
5. © 2026 Thoughtworks Context Engineering: Beyond MCP servers and rules files 5 Skills Rules Specs Subagents Plugins Commands MCP Servers
6. © 2026 Thoughtworks Skills example Modularisation of rules Can be loaded by LLM just-in-time Can include more files Often refer to already installed CLIs
7. © 2026 Thoughtworks Context Engineering “in a nutshell” Reusable instructions and conventions Skills Rules Specs Com- mands Context interfaces MCP Servers Tools Skills “Intelligentlyˮ loaded, just in time Manage and monitor context size
8. © 2026 Thoughtworks
9. © 2026 Thoughtworks Subagents Main conversation (“orchestratorˮ) Research Implemen- tation Code Review spawn report spawn report spawn report
10. © 2026 Thoughtworks Context Engineering: Ask yourself… Coding conventions you want to amplify? → Skills Workflows to build for modernisation initiatives? → Subagents, Skills Tools that should be available in your org? → CLIs, MCP servers, LSPs, … Versioning and distribution? Is it making things better, or worse? What are practices you want to amplify? → Skills
11. © 2026 Thoughtworks More agent autonomy, less human supervision
12. © 2026 Thoughtworks Supervised Unsupervised
13. © 2026 Thoughtworks
14. © 2026 Thoughtworks CLIs 14
15. © 2026 Thoughtworks Headless → Integratable into pipelines
16. © 2026 Thoughtworks TUI vs GUI
17. © 2026 Thoughtworks A familiar beast rears its head: Dev Sandboxes “X is not installedˮ OutOfMemoryError Internet access: Yes or no, when and where to?
18. © 2026 Thoughtworks Intermezzo: The hype du jour
19. © 2026 Thoughtworks From: “Welcome to Gas Townˮ https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16dd04
20. © 2026 Thoughtworks Agent Swarms
21. © 2026 Thoughtworks Claude Code: Agent teams
22. © 2026 Thoughtworks Back to the present…
23. © 2026 Thoughtworks Less supervision: Ask yourself… Where do you want to experiment with cloud agents? How do you help your teams gauge the appropriate level of supervision?
24. © 2026 Thoughtworks Risk assessment
25. © 2026 Thoughtworks 25 Probability …that AI gets something wrong Impact …if AI gets something wrong Detectability …that AI got something wrong → Know the AI tool, know and engineer the context → Know your confidence level in the requirements → Know the use case criticality → Know your feedback loops Which workflow? How much review? How long without supervision?
26. © 2026 Thoughtworks 26 Probability …that AI gets something wrong Impact …if AI gets something wrong Detectability …that AI got something wrong “You have to be this tallˮ to reduce supervision
27. © 2025 Thoughtworks More autonomy, less supervision: Beware of security and cost
28. © 2026 Thoughtworks Security - unwanted command execution > Prompt Injection > Bypassing allow lists
29. © 2026 Thoughtworks Security - extraction of secrets https://grith.ai/blog/clinejection-when-your-ai-tool-installs-another
30. © 2026 Thoughtworks The lethal trifecta Access to Private Data Ability to Externally Communicate Exposure to Untrusted Content Simon Willison https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/
31. © 2025 Thoughtworks Cost - “the honeymoon is over” Beginning of 2024 “Generating 100 lines of code only costs about 12 cents, compare that to a developer salary!ˮ Keynote at Craft Conf 2024 (simplified) Summer 2025 viberank.app
32. © 2025 Thoughtworks Cost - “the honeymoon is over” Beginning of 2024 “Generating 100 lines of code only costs about 12 cents, compare that to a developer salary!ˮ Keynote at Craft Conf 2024 (simplified) Summer 2025 viberank.app $20 Flat Rates $200 Flat Rates Request limiting
33. © 2025 Thoughtworks https://bsky.app/profile/edzitron.com/post/3metzkqobsk2z
34. © 2025 Thoughtworks Why do 100 LOC cost more like $2,50, not $0,12? Plan Review the plan Research existing code Implement the first task Make changes Run the tests Fix the tests Check lint errors Fix lint errors Check the browser “Itʼs not rendering, debugˮ Fix Code review Improve a method Summarise
35. © 2026 Thoughtworks Less human supervision: Ask yourself… How to sandbox coding agents? Does everybody understand the lethal trifecta? Where do you want to experiment with cloud agents? How do you help your teams gauge the appropriate level of supervision?
36. © 2025 Thoughtworks More autonomy, less supervision: What about maintainability?
37. © 2025 Thoughtworks Context engineering Architectural constraints Custom linters, structural tests “Garbage collectionˮ
38. © 2025 Thoughtworks Experiment: Structural tests as agent feedback “External SDKs may only be imported by files in server/clientsˮ
39. © 2025 Thoughtworks A mental model
40. © 2025 Thoughtworks Harness Executed by GPU inference Executed by CPU Coding Agent Principles Rules, Examples Ref Docs How-tos CfRs Guides feed-forward Runtime state Static state Agent-as- judge Sensors Mutation testing feed-back self-correction CPU-executed when possible Shift sensors left Human Steering Loop Uses to build steering loop LSPs CLIs, scripts
41. © 2026 Thoughtworks 41 Probability …that AI gets something wrong Impact …if AI gets something wrong Detectability …that AI got something wrong Harness Our trust in the agent
42. © 2025 Thoughtworks Still LOTS of open questions about verifying functionality
43. © 2025 Thoughtworks Maintainability Operability Behaviour
44. © 2025 Thoughtworks Data dashboard, Typescript Will harnessable topologies be the new abstraction layer? Structure Tech Stack Harness template instantiate Guides Sensors
45. © 2025 Thoughtworks In summary…
46. © 2026 Thoughtworks Strong forces are tempting humans out of the loop Where can your organisation give in to that pull, where is it dangerous? Context engineering Powerful lever of amplification – both good and bad
47. © 2026 Thoughtworks Mehr: birgitta.info martinfowler.com/articles/ exploring-gen-ai.html
48. © 2026 Thoughtworks Thank you