Available prototype
Coding agents.
Source-backed observations, repository reads, and evidence carried between sessions. Local validation keeps reuse grounded in the current code.
A shared execution layer for AI agents
Good work.
Worth reusing.
Your next agent shouldn’t have to start over.
Let useful work carry forward.
Starting with code. Built for what comes next.
Work once. Carry it forward.
⌘
Agent 01Discovers the source
✓⌁
Agent 02Builds on the evidence
✓ Source verified
↳
Agent 03Picks up where you left off
Evidence, with its context.
Keep your agents.
Give their work continuity.
Give their work continuity.
Vlaid works alongside existing coding agents.
Same tools. Shared evidence. Less rediscovery.
01 / The idea
A second agent.
A head start.
Useful work shouldn’t disappear when a session ends. Vlaid preserves source-backed evidence so the next agent has something to build on.
The evidence lifecycleIllustrative coding workflow
Source observationCaptured
src/payments/retry.pyL18–21
18
def should_retry(response):19
if response.status == 429:20
return True21
return response.status >= 500SourceExact file excerpt
ProvenanceRead result + source hash
ScopeThis repository
Recorded evidence, not a generated answer.
No model required for deterministic selection.How the evidence works
See it in practice
Less rediscovery.
Same task.
A recorded coding comparison, from the first read to the final check.
02 / Measured progress
Evidence before
big promises.
Measured across coding and operations workflows. We keep the full experiment history, including earlier versions and results that didn’t improve.
Read the benchmark notes35.1%
Lower cost on operations tasks
Across all valid paired operations runs.
Earlier and reverted versions included.
20.3%
Lower cost on coding tasks
72 out-of-sample coding pairs.
95% confidence interval: 15.8–24.6%.
1,827
Valid operations pairs
Reported across October 17–19.
Every included session passed.
Internal benchmarks, not a savings guarantee. Figures measure API-equivalent cost, not subscription bills. Workloads differ; coding results include Claude prompt-cache settings. Methodology and limitations ↗
03 / A longer view
Start with code.
Follow the work.
Repeated discovery happens wherever agents gather information. We’re starting with a problem we can inspect and verify.
Research direction
Every team’s agents.
A future where research, finance, and operations can build on work already gathered. Shared sources. Clear provenance. Context that carries forward.
Concept in development. Not currently supported.Build on what
came before.
Infrastructure for agents that work together.