WEFT IN PRACTICE

See what your agents can take off your plate.

More work in parallel. Fewer repeated explanations. Checks before acceptance. A model budget that goes further.

01 / WORK IN PARALLEL

Start another task while the first one runs.

Keep talking to your agent. It launches and manages the workers, tracks their progress, and brings you results and questions. You have one conversation to follow, with less to keep in your head.

Independent tasks run in separate worktrees with their required checks. If one needs a fix, your agent manages the revision while unrelated work continues.

From the work record

In a recorded repository migration, three agents launched within 26.6 seconds and returned task-completion responses within about 40 minutes. Two changed code; one confirmed no change was needed. Later checks found remaining work and held a dependent step until it was addressed.

That work used Weft with the Cloud plugin. The illustration shows how to apply the coordination pattern with local OSS tasks and workers.

TWO TASKS / ONE WAY TO CHECK THE WORK
YOUR REQUEST“Add search. Fix the export.”
AGENT 01SearchBuild + tests
AGENT 02CSV exportBuild + tests
The export test finds missing rows.That task waits. The agent fixes it and checks again.
Both results ready to inspect.Each task passes its required checks.

Illustrative local workflow · independent tasks can run together.

Inspect the results after their checks pass.
Follow the agent conversation
AGENT / WORK RECORDOSS · local tasks + your agent
  1. You
    Move the migration forward. Keep independent work in parallel.
  2. Your agent → Weft
    Record three tasks, their acceptance criteria and dependencies in the local project.
  3. Your agent
    Delegate separate workstreams to agents in separate worktrees.
  4. Worker agents → Weft
    Two changes complete. The third task confirms no change is needed. Results recorded.
  5. Repository check
    Remaining consumers found. Hold the dependent removal step.
  6. Your agent
    Assign the correction. Recheck consumers before continuing.
Illustrative OSS replay, adapted from recorded parallel work. Playback changes this example only.
Source and scope

The recorded June 2 migration launched three agents within 26.6 seconds and received their completion responses in about 40 minutes. One lane was a verified no-op. A later correction migrated 70 consumers; the final sweep found and migrated seven more. Claude launched the agents and applied the consumer check. This replay translates that practice to local OSS records; it is not a recording of native worker dispatch.

02 / KEEP THE ANSWER

A new session shouldn’t mean explaining everything again.

Have your agent search project memory before making a significant choice. When the answer is missing, it asks you and records your decision and the reason.

A later agent can retrieve it before continuing. The record stays with your local project when you change agents or models.

From the work record

A recorded session searched for prior guidance, found no answer, offered three options, and saved the human’s choice. Related decisions were retrieved by later workers. The export-date example makes that pattern concrete.

A QUESTION TODAY / A REASON TOMORROW
AGENT → YOU“Keep the original dates?”No prior project decision found.
YOU → AGENT“Yes. The dates matter.”
RECORDED IN LOCAL PROJECT MEMORY
Decision
Keep original timestamps.
Reason
The export must preserve the source history.
A LATER AGENT SEARCHES THE RECORD“Found the answer—and why.”

Illustrative handoff · agents record and retrieve decisions.

Retrieve the answer for later work.
Follow the agent conversation
AGENT / WORK RECORDOSS · decisions in your local project
  1. Your agent
    The importer could change existing dates. That changes the meaning of the records.
  2. Your agent → Weft
    Search prior decisions about import timestamps.
  3. Local decision records
    No matching answer found.
  4. Your agent → you
    Keep the original dates, use the import date, or ask before changing them?
  5. You
    Keep the original timestamps.
  6. Your agent → Weft
    Record the decision and its reason in the local project. Keep the original timestamps.Reason: People use dates to understand the order of their records.
  7. A later agent session
    Retrieve the decision before continuing the importer. Keep the original timestamps.
Illustrative replay, based on recorded agent workflows. Playback changes this example only.
Source and scope

Your agent follows the search → ask → record → retrieve policy using Weft's tools. A recorded June 18 episode includes an empty search, three options, a human answer and a saved decision. Related decisions were retrieved and applied by later workers. The importer here is an illustrative example, not a literal transcript. The decision stays in the local project; sharing it across separate machines is a Cloud capability.

03 / FIX BEFORE ACCEPTING

“Done” still has to pass your checks.

The agent finishes an export. A test finds that it changed the original dates. With that check required by the workflow, the task waits for a fix.

The agent corrects the behavior and runs the checks again. You can inspect the result and its verification before accepting it.

This is an illustrative repository test. OSS runs your configured checks; Cloud AI review is an optional addition.

ONE TASK / A RECORD TO CONTINUE FROM
TASK 024Ready for acceptance

Add a CSV export.

Owner
Implementing agent
Done when
Export passes the repository’s checks
  1. Scope recordedOne bounded change.
  2. Check → revisionThe failed check sends work back.
  3. Build + tests passThe result is ready for acceptance.
NEXT TASKReuse the workflow. Consult the decisions.

Illustrative local task · your agent does the work.

Start the next task with context.
Follow the agent conversation
AGENT / WORK RECORDOSS · your repository's build and tests
  1. Implementing agent
    Importer implemented. Run the required build and test verification.
  2. Repository test
    FAIL · importing an entry changes its original timestamp.
  3. Workflow
    Verification has not passed. This transition is held.
  4. Implementing agent
    Preserve the source timestamp. Add a regression test.
  5. Repository checks
    PASS · build and tests. Verification evidence recorded.
  6. Your agent
    The required checks passed. Inspect the result before merging.
Illustrative configured local verification workflow. Playback changes this example only.
Source and scope

OSS supplies local workflow stages, structural checks and build/test verification. The test shown is an example of a repository test you supply. This is not hosted AI review, and a passing test does not replace your acceptance judgment.

04 / CHOOSE THE MODEL

Spend the stronger model’s effort where it helps.

A small, well-defined change can go to a lower-cost model. Keep a stronger agent on supervision, with the same task, acceptance criteria, and verification results.

You can try different assignments while reusing the process.

From the work record

A recorded task paired a Pi implementer using DeepSeek Flash with a Claude Opus supervisor. The implementation turn took 26.905 seconds and reported an estimated cost of $0.008538006. That is the implementation turn alone; supervision, review, and tests were additional work.

The supervising agent chose the models. The verified change merged as PR #1195.

A BOUNDED CHANGE / TWO DIFFERENT JOBS
IMPLEMENTLower-cost model

Make the scoped change using the task and acceptance criteria.

SUPERVISEStronger model

Inspect the patch, challenge the approach, and run the checks.

The same acceptance criteria and required checks.

You or your agent choose the models. Use your own accounts.

The same acceptance criteria and checks apply.
Follow the agent conversation
AGENT / WORK RECORDYour agent · your model accounts
  1. Your agent
    Small, bounded task: remove a duplicate reachability check.
  2. Weft work record
    Task, acceptance criteria and the allowed files are available to both agents.
  3. Pi · DeepSeek Flash
    Implement the functional change.
  4. Claude Opus · supervisor
    Check the premise in the source. Demonstrate the regression check fails before the fix and passes afterward.
  5. Supervisor
    Run the full repository checks. Inspect the patch.
  6. Work record
    Verified change merged as PR #1195.
Adapted from a recorded implementation and supervision workflow. Playback changes this example only.
Source and scope

One recorded Pi implementation turn took 26.905 seconds, with a model-reported cost estimate of $0.008538006. That excludes supervision, review, tests and CI. The supervising agent selected the models; this is not automatic model-to-complexity classification by Weft. This pattern uses your own coding engines and model accounts, with no Cloud requirement.

GO DEEPER

Inspect the evidence behind the examples.

The development study follows real agent work over time. These examples translate useful patterns into local OSS workflows.

Read the development study ↗Read the Trust Topology research ↗Start with a small task →