Blog
Engineering insights on building AI agents that learn.
HyperSTRUCKSep 30, 2026
Holding conflicting claims until they are verified
Measured: 24 of 24 payment attributes raised to consequential in replay
Latest · 5 min read · Thomas Lynch
How we made agents hold a claim instead of believing it or throwing it away
Some things an agent is told should not be acted on yet, and should not be thrown away either. Hyperstruck holds them with a reason, tells the agent a change is waiting, and releases them once the reason stops applying.
HyperSTRUCKSep 10, 2026
Tracing every record to the document version it came from
Measured: reverse lookup took 5 milliseconds against 41 with a dedicated table
Sep 10, 2026 · 4 min read
How we made every record carry its own justification
Most systems can name the file a record came from. They can't tell you if that file has since changed, or been deleted. We built a document model that can, down to the exact sentence.
HyperSTRUCKAug 23, 2026
Carrying learnings across renamed tools by matching on function
Measured: 56% carried across on the hardest test, none on the wrong tool
Aug 23, 2026 · 2 min read
Experience Earned Through Tools Is Lost When the Tools Are Renamed
An agent that works something out writes it down in the names of its own tools. Rename them and the learning is gone, unless something matches on what the tools actually do.
Earlier posts
25 posts since May 2026Aug 5, 2026 · 6 minAug 5, 2026Replit, OpenAI and Meta Fixed Agent Context. Not One of Them Published a Number.6 min
Measuring whether context fixes make agent answers correct
Jul 20, 2026 · 12 minJul 20, 2026You Have an Agentic Mandate. Now What?Tony Truong12 min
Seeding an experience layer, then designing agent architecture on it
Jul 19, 2026 · 3 minJul 19, 2026You Already Wrote the Rule Down3 min
Why a rules file does not stop repeat mistakes
Jul 19, 2026 · 3 minJul 19, 2026Plausible Is Not Proven3 min
Choosing approaches by what has held up in your environment
Jul 19, 2026 · 3 minJul 19, 2026Nobody Wrote Down the Order3 min
Learning the order operations have to run in
Jul 11, 2026 · 5 minJul 11, 2026Agents Can Now Start With ExperienceTony Truong5 min
Preloaded domain experience for cloneable agents
Jul 10, 2026 · 4 minJul 10, 2026Stopping Agents From Reporting False Success4 min
Checking agent actions and success claims against what tools returned
Jun 23, 2026 · 3 minJun 23, 2026Innate Intelligence3 min
Shipping agents with generic invariants earned across prior work
Jun 19, 2026 · 10 minJun 19, 2026Run Your SDLC in Co-Pilot Mode: A Developer's Getting-Started GuideTony Truong10 min
Wiring recall and capture hooks into each software development stage
Jun 9, 2026 · 3 minJun 9, 2026A Hunch Is Not a Rule3 min
Scoring learnings separately on usefulness and how well established
Jun 8, 2026 · 2 minJun 8, 2026Where Agents Lose the Plot2 min
Writing final answers that address the goal and match the run
Jun 3, 2026 · 4 minJun 3, 2026Sharing Learnings Without Leaking Secrets4 min
Classifying sensitive values from tool declarations before sharing learnings
Jun 2, 2026 · 4 minJun 2, 2026Why we didn't build Hyperstruck on a graph4 min
Building agents from a goal instead of an authored graph
May 30, 2026 · 2 minMay 30, 2026A Repeated Lesson Is Not a Duplicate2 min
Treating repeated learnings as corroboration instead of duplicates
May 29, 2026 · 2 minMay 29, 2026Agents Need to Forget2 min
Retiring learnings on contradicting evidence instead of time decay
May 27, 2026 · 3 minMay 27, 2026Labels Are a Liability3 min
Replacing write-time category labels with stored execution evidence
May 23, 2026 · 3 minMay 23, 2026Most of What Your Agent Learns Is Garbage3 min
Filtering tautological learnings and keeping prescriptive ones
May 21, 2026 · 3 minMay 21, 2026What Are Learnings?3 min
Defining learnings as behavior changes with a trigger condition
May 21, 2026 · 4 minMay 21, 2026Frontier Performance from an Open-Source Model4 min
Lifting an open-source model's SWE-bench Verified score with an agent
May 19, 2026 · 5 minMay 19, 2026The God Model Trap5 min
Pinning each component to a benchmarked model for easy swaps
May 19, 2026 · 4 minMay 19, 2026LLMs Can't Count to Five4 min
Scoring with extraction instead of rating scales
May 15, 2026 · 3 minMay 15, 2026Your Agents Start From Zero Every MorningTony Truong3 min
Retrieving prior learnings to improve a coding agent's plan
Product updates
What Hyperstruck has shipped lately
What changed for customers and how the platform is running, one issue at a time.
All product updates