Version control for work in motion.
Keep building, exploring, and returning to the moments that matter.
Git records the states we choose to publish. DeltaGit preserves the work that leads to them.
Every meaningful operation can become a lightweight checkpoint. Agents can branch from earlier moments, explore alternatives, compare outcomes, recover exact workspace states, and promote only the path that matters.
flowchart LR
C0["Checkpoint C₀"] --> O1["tool operation"]
O1 --> C1["Checkpoint C₁"]
C1 --> OA["try approach A"]
C1 --> OB["try approach B"]
C1 --> OC["try approach C"]
OA --> A["Candidate A"]
OB --> B["Candidate B"]
OC --> C["Candidate C"]
A --> E["Compare outcomes"]
B --> E
C --> E
E --> P["Promote selected state"]
P --> G["Ordinary Git"]
classDef checkpoint fill:#e8f1ff,stroke:#2563eb,color:#172554;
classDef candidate fill:#f5f3ff,stroke:#7c3aed,color:#2e1065;
classDef decision fill:#ecfdf5,stroke:#059669,color:#022c22;
classDef git fill:#fff7ed,stroke:#ea580c,color:#431407;
class C0,C1 checkpoint;
class A,B,C candidate;
class E,P decision;
class G git;
DeltaGit is designed for a future where software agents learn not only from final answers, but from the full trajectory of attempts, failures, corrections, and discoveries.
flowchart LR
S["Select a prior checkpoint"] --> X["Explore possible paths"]
X --> T["Retain trajectories"]
T --> V["Evaluate outcomes"]
V --> L["Learn what worked"]
L --> N["Improve the next generation"]
N --> S
V --> R["Reject and recover"]
R --> S
classDef state fill:#e8f1ff,stroke:#2563eb,color:#172554;
classDef explore fill:#f5f3ff,stroke:#7c3aed,color:#2e1065;
classDef evaluate fill:#fff7ed,stroke:#ea580c,color:#431407;
classDef improve fill:#ecfdf5,stroke:#059669,color:#022c22;
classDef reject fill:#fef2f2,stroke:#dc2626,color:#450a0a;
class S state;
class X,T explore;
class V evaluate;
class L,N improve;
class R reject;
DeltaGit provides the durable state and trajectory history for agentic search, reinforcement learning, and recursive improvement—while keeping Git as the collaboration and delivery layer.
Checkpoint freely. Explore broadly. Learn continuously. Promote deliberately.