Injected context saves 25% in time and tokens.
FML injects compact task context from prior sessions, repo history, docs, tests, incidents, and product signals before the agent starts exploring.
Context injection
The bundle selected for one checkout agent run.
Prior sessions
17
checkout retry, auth migration, PCI review
Related PRs
5
merged fixes, reverted retry path, review comments
Docs and tests
8
architecture notes, runbooks, relevant test files
Production signals
4
Sentry incidents and PostHog events tied to the task
Session starts
Prompt: "fix checkout retry bug without changing PCI boundary"
FML retrieves context
prior sessions, related PRs, docs, tests, Sentry, PostHog, and team rules
Agent starts with the bundle
less repo-crawling, fewer clarification turns, fewer repeated mistakes
25%
faster in replay benchmarks, with fewer discovery turns and lower token spend
What changes when the agent starts with history
FML context injection completed comparable tasks about 25% faster with fewer turns.
Cold start
Agent explores from scratch
With FML context
Agent starts with the useful bits
About 25% faster on comparable tasks, with fewer turns and lower token spend.
How we measure it
The 25% figure comes from replay benchmarks on real work, not synthetic demos. Results vary by repo and task mix; 25% is what we see across the benchmark set.
Replay real tasks
Each benchmark task is a prompt from a real session, replayed in the same harness against the same repo state.
Run both conditions
One run starts cold. The other starts with the FML context bundle. Same model, same tools, same prompt.
Compare the runs
We compare time to completion, agent turns, and token spend. With injection, comparable tasks finished about 25% faster.
What gets injected
Small bundles of task-specific history, not a giant dump of everything the company knows.