Self-evolving agents · measured on our own team first

The hundredth time should be faster.

re-forge builds self-evolving agents on your team's compounding memory: solve a problem once and every agent gets sharper. No one re-derives a fix, no one collides mid-build, and the hundredth run knows what the first ninety-nine learned.

Join the waitlist

View on GitHub

how it compounds

Capture · Detect · Evolve

capture

installed by a human, consented, kill-switched

116k+ messages 29 repos

detect

every session read for repeats

200 repeated procedures [detected]

evolve

agents sharpen; every gain measured and labeled

up to 123 h [upper bound] $90.5k reused [measured]

our own 2-person team, Jun 20 – Jul 2 2026

The instrument

Your team's memory, on one screen.

The dashboard shows the memory at work: what your agents already know, where work repeats, every number labeled.

Hours reclaimable · do it once

123h

Up to · upper bound

$90.5kReused

reusedcache reads

input served from cache · all time

Repo graph · live592 nodes · 579 edges · live scan

200 repeated procedures· up to 123 h reclaimable· $90.5k reused

$90.5k = 6.54B tokens served from cache × ~$13.8/MTok (fresh list price − cached price) 123 h = 1,136 repeats × ~6.5 min each · only the repeats count, never the first run

labels travel with every number your repo, as a living map

The team

Parallel work, without the pileup.

Two people, code that overlaps. PreFlight sees it before either build does: prompts become specs, specs get checked against the live repo graph, boundaries get pinned, both changes land clean.

akash · terminal

$ task: add due-date reminders

· drafting spec…

$ task: add due-date reminders

· checking repo graph…

$ task: add due-date reminders

· boundaries pinned

$ task: add due-date reminders

↺ context: 2 specs · 3 pinned files

$ done: reminders shipped, list untouched

preflightmerged · 0 conflicts

preflight · listening

akash

intent: reminders on due dates
expects: tasks/models.py, tasks/views.py
interface: TaskItem

shine

intent: overhaul the task list UI
expects: templates/list.html, tasks/views.py
interface: TaskItem

tasks/models.py · tasks/views.py · templates/list.html

akash

intent: reminders on due dates
+ boundary: reminders never rewrite list rendering
+ interface: extend TaskItem, do not modify

shine

intent: overhaul the task list UI
+ owner: shine holds list.html, akash consumes
+ interface: extend TaskItem, do not modify

measured on our Team Sync demo: 13 predicted conflicts became 0

shine · terminal

$ task: overhaul the task list UI

· drafting spec…

$ task: overhaul the task list UI

· checking repo graph…

$ task: overhaul the task list UI

· boundaries pinned

$ task: overhaul the task list UI

↺ context: 2 specs · 3 pinned files

$ done: list UI shipped, no conflicts

Trust

Read every byte before you trust it.

$ cat CAPTURE-MANIFEST.json
{ "reads":  ["~/.claude/projects/**/*.jsonl", …],
  "sends":  ["prompts", "answers", "tool results (clamped)"],
  "never":  ["API keys", "secret-denylist files"],
  "kill_switch": "~/.reforge/DISABLED" }

$ reforge off   # capture stops, immediately

Open-source client, read every line.
Nothing captured before a human consents.
One file kills it all: ~/.reforge/DISABLED

The labels

Words we can't take back.

Measured

Observed in real session logs

Detected

Found by clustering, count it yourself

Upper bound

A ceiling, never a promise

Reused

Already happened, never double-counted

Founders

Shine Gupta CEO S Akash CTO

reach us through the waitlist form

Give your team's agents a memory.

One key signs this device in.

Join the waitlist

View on GitHub