Sarvam AI
Sarvam Motif

Coding Agents

Better code in the long run

Coding Agents work across the codebase and saves checkpoints as it goes. The plan stays visible, so you can change direction while the run is live.

user@sarvam:~/acme/payments-service$ sarvam-code

Coding Agents (v0.38.0)

directory: ~/acme/payments-service

Migrate the payments data layer

Move the payments service off the legacy ORM without downtime. Keep the public API stable. Stop the run if a contract test fails.
  1. Thought for 6 seconds
  2. Read billing/repository.py, billing/models.py, tests/contract/
  3. Found 214 uses of 'session.execute'

  4. I'll add a compatibility shim, migrate one module at a time, and run the contract suite after each change. Billing first.

  5. billing/repository.py +shim +41 -0
  6. billing/models.py +data layer +96 -12

  7. Contract tests pass. 3 of 5 steps done. Moving to the ledger module now.

Move on to the ledger module
Try Coding Agents

A visible plan for work across the codebase.

Coding Agents break the job into steps, writes and tests each change, and saves checkpoints as it goes. The plan stays visible throughout the run.

Work across the codebase

Give Coding Agents a refactor, migration, or performance problem that crosses services. It works through one part at a time while the original goal stays on the plan.

Migrate 214 database calls

Replace the legacy ORM across the payments service, one module at a time.
  1. Thought for 4 seconds
  2. Read 214 call sites across 18 modules
  3. billing/repository.py +41 -0
  4. ledger/queries.py +63 -8

Steer the run

The whole plan stays on screen. Change the goal, move a step, or stop the current approach. Coding Agents pick up from the work already done.

Reorder the migration plan

Skip the webhook rewrite for now and finish the ledger module first. Keep the rest of the plan.
  1. Thought for 2 seconds
  2. Dropped branch 'reconnect-webhooks'

  3. Plan updated. The ledger module is step 3. The webhook rewrite is last.

Pick up from a checkpoint

When a test fails or a path reaches a dead end, Coding Agents returns to a checkpoint and tries again. Completed steps stay in place.

Recover from a failed test

The ledger contract tests failed. Resume from the last checkpoint and keep the completed steps.
  1. Thought for 7 seconds
  2. Read tests/contract/test_ledger.py
  3. ledger/queries.py fix decimal rounding +6 -4

  4. Contract suite passes. I resumed from the last checkpoint and kept the completed work.

Use expensive models only for the tasks that need them.

Pick a model for each task

A refactor and a rollback review call for different models. Coding Agents routes them separately during the same run.

user@sarvam:~/acme/payments-service$ sarvam-code

Coding Agents (v0.38.0)

directory: ~/acme/payments-service

Route each task

Write the data-layer boilerplate, then review the rollback path. Route each one to whichever model suits it.
  1. Classified 2 task types in this run
  2. Scaffolding routed to a fast, lower-cost model
  3. Rollback review routed to a reasoning model

  4. Scaffolding and rollback review went to different models. Coding Agents routed each task separately.

Show the cost split by task

Train a model for a specialised task.

Coding Agents can train a task-specific model during the run using your data and requirements, then apply it to the job. For underwriting, that can mean training on your risk parameters.

user@sarvam:~/acme/underwriting$ sarvam-code

Coding Agents (v0.38.0)

directory: ~/acme/underwriting

Train an underwriting model

Train a task-specific model on our portfolio and risk parameters, then use it for this pricing task.
  1. Thought for 9 seconds
  2. Read risk/params.yaml, historical policies
  3. Prepared evaluation holdout

  4. Training finished this run. The model is ready for your review on this risk band.

  5. risk/underwriting.model ready for review

Frequently Asked Questions

Put Coding Agents to work on a real codebase.

Talk to us