Execute through a harness
Give the system a goal, repository, tools, and a bounded budget. AO is the intended worktree and session plane; the adapter keeps that boundary explicit.
Open-source infrastructure for agent fleets
Agents can finish a task.
Forge keeps the proof.
Forge is the evidence-gated learning layer above your agent harnesses. It verifies the artifact, keeps failure context, and turns a useful repair into a candidate lesson—without confusing reflection with learning.
local-first ledger + verifier
bounded repair budgets
honest promotion gates
static illustrationnot a live tracker
01 / the product
Experience should compound.
Proof makes it portable.
Most agent systems optimize the next turn. Forge is concerned with the turn after that: what can be safely reused, what should be rejected, and what evidence supports the decision.
It sits above execution. AO or another harness may run the worker; Forge owns the acceptance check, repair budget, evidence ledger, and learning gate.
02 / how it works
A loop with a stopping rule.
No success message counts without an artifact and an independent check.
worker boundary
Verify the artifact
Run frozen or independent acceptance checks in a verifier-owned context. A worker’s final message is evidence of communication, not proof of completion.
Repair once, then stop
Classify the failure and allow a bounded repair path. Stuck, hidden-blocked, or no-op sessions are nudged once and terminated rather than looped forever.
Reflect, then earn reuse
Reflection can propose a candidate skill from bounded evidence. Forge still requires applicability, A/B benefit, and held-out no-regression checks before promotion.
03 / observed artifact
A real run, not a dashboard mockup.
The current Forge submission artifact starts red, repairs once, and ends green on a frozen C0 task. GPT-5 Nano returns a structured candidate skill. The candidate remains a candidate.
baselineexit 1 · 4 failed
repair1 bounded attempt
finalexit 0 · 4 passed
regression61 local tests
observed local artifactcandidate ≠ promoted
04 / local proof demo
Read the change.
This is a small local illustration of the evidence contract. Switch states to see the same artifact move from failure to bounded repair to a verifier-backed result. It does not poll AO or represent a live run.
local illustration · no backend · no polling
FORGE / LOCAL ARTIFACTbaseline / failed
Frozen verifier found the break.
EXIT 14 frozen checks failed
baseline evidencestatic / illustrative
04 / the boundary
Say what is true.
Leave the rest visible.
Current product boundaryread before you run
AO health, readiness, and isolated worktree creation are observed. A reliable artifact-producing AO completion lifecycle is still an open proof point. Cross-harness transfer and Supermemory remain future integrations.
05 / get started
Install the runtime.
Inspect the proof.
Forge is currently a local-first Python project. The public site is static; the learning loop and evidence ledger run in your checkout.
Full installation guideshell
# clone and install
git clone https://github.com/LangerSword/forge.git
cd forge
uv syncverify the checkout
uv run forge status
uv run pytest -q