obversa
Read the docs

Run your software company. Don't just code in it.

A process runtime for software teams. Agents, people and reviews share one durable process and one record.

Read the docs See how it works
feature-dev recorded
  1. spec started
  2. spec completed Let people export a report with a header and all its rows.
  3. approve started
  4. approve waiting Waiting for human approval.
  5. approve approved simulated
  6. approve completed Implement the report export
  7. implement started
  8. implement completed report.csv
  9. verify started
  10. verify completed The exported report has the expected header and rows.
  11. review started
  12. review completed correctness, safety, clarity, scope
  13. proof started
  14. proof completed verification.json
  15. release started
  16. release completed Ready to release
A feature ticket becomes a specification and waits for a simulated human approval. The implementation produces a report export, verification checks its exact content, and a review checks correctness, safety, clarity, and scope. The passing checks are saved as a proof artifact, and the report is marked ready to release.
One system, not a scatter.
The run is one process with one record, not a pile of chats, terminals, tools and tribal notes.
Judgment where it matters, friction where it doesn't.
Agents do the work. Applications can connect stored callback requests and local review answers to the direct executor's pause and resume API. The local review surface is what ships today.
A crash does not lose the run.
A fresh executor recovers recorded progress; an uncertain attempt pauses unless retry was declared safe.
Audit for free.
The event log records run transitions and links to stored attempt and review evidence.

How it works

  1. describe

    Write the work as a graph.

    Steps, dependencies and gates share one graph contract.

  2. run

    The runtime runs bounded node attempts.

    Each attempt has a budget. Writer and reviewer roles use separate lanes. The review loop can require a diverse quorum of reviewers.

  3. record

    Every step lands in the event log.

    State comes from the log. A fresh executor recovers recorded progress; an uncertain attempt pauses unless retry was declared safe. The direct executor records a request when an application asks it to wait. The application connects human answers and resumes the run.

The workflow and its record
16 of 27 stored events
workflow callback gate SpecItem, Quantity Implementreport.csv Verifypassed Review4 checks passed Proofverification.json ReleaseReady to release THE RECORD · 16 OF 27 STORED EVENTS +74 msspec started +88 msspec completed +112 msapprove started +125 msapprove waiting for a person +179 msapprove answered +212 msapprove completed +238 msimplement started +250 msimplement completed +278 msverify started +289 msverify completed +313 msreview started +327 msreview completed +356 msproof started +396 msproof completed +426 msrelease started +466 msrelease completed
Sixteen of twenty-seven stored events, at their recorded offsets.

Pull-request review loop

A pull request receives review feedback, fixes missing input validation in one revision, and becomes ready to merge.

The review loop and its record
10 of 16 stored events
workflow, a review loop revision 1 of 1 DraftValidate empty input Reviewthree checks PublishReady to merge THE RECORD · 10 OF 16 STORED EVENTS +54 msdraft started +64 msdraft completed +85 msreview started +96 msreview completed +119 msdraft started +130 msdraft completed +153 msreview started +166 msreview completed +191 mspublish started +203 mspublish completed
Ten of sixteen stored events, at their recorded offsets.

The four layers of an agent system

The picture the market already uses, with Obversa placed on it.

Loopa node attempt
A bounded node attempt. Writer and reviewer roles use separate lanes. The review loop can require a diverse quorum of reviewers.
Graphthe graph contract and its forms
Pipelines and review loops use one graph contract. State comes from the event log. A fresh executor recovers recorded progress; an uncertain attempt pauses unless retry was declared safe.
Harnessengine plugins
We don't replace your harness. Six plugins drive the ones you already run: Claude Code, Codex, Grok, OpenCode, the Anthropic API and the Agent SDK, with recorded outcomes and links to stored evidence.
Meta-harnessworkflows over agents and people
The layer this runtime is for: agents and people in one process. Company-scale orchestration is later.

Bring the harness you already run.

A larger model where judgement matters. A smaller one where the job is narrow. One record over the mix.

Choose a supported harness and model pairing for each step.

What's real today

Shipped

  • The pipeline form: ordered stages on the durable graph executor.
  • The review loop: separate writer and reviewer lanes, optional reviewer diversity, and callback support.
  • Crash-safe resume: a fresh executor recovers recorded progress; an uncertain attempt pauses unless retry was declared safe.
  • Six engine plugins, and the local review surface.

On the roadmap

  • Remote, mobile and third-party review surfaces.
  • Company-scale orchestration across many processes.
  • Query and export tools for the audit trail. Run events and stored artifacts provide the source for audit queries.

Run the work. The record comes with it.

Read the docs Source on GitHub