Skip to content

Thesis

Why Assent should exist, and what would prove it shouldn’t.

Agents write more of the code, but the decisions still belong to the teams that own it. Assent is designed to put those decisions first, before the run.

Working directioncommercial thesis unvalidatedAdopted 2026-10-06

An illustration of the thesis. Nothing here is measured.

Concept UI · illustrative data

What this shows

  1. Agents write code fast: Reservation API, Expiry serializer, Checkout expiry
  2. Waiting on a decision: Duration or absolute timestamp?, Persist expiry across restarts?, Clock behavior on restart.
  3. Owners still decide one at a time: Reservations, Checkout
  4. With Assent, questions go to owners first

Fig. 1An illustration of the thesis. Nothing here is measured.

What we believe, why, and how we’ll know if we’re wrong. No external evidence supports this yet; it is the direction we are testing.

Design noteThis page describes how Assent is designed to work, and why. Intent, not results.

The question Assent should answer.

Was what we decided implemented correctly, what was verified and what was not, and does anything need my decision?

Engineering teams should decide how their software behaves, and Assent should run the coding agents that implement those decisions.

Concept UI · illustrative data
Pull request ready for review#218
Expose reservation expiry to Checkout
9c2e71dmainevidence attached
Checked
  • Path scope
  • Approved-contract comparison
  • Consumer contract tests
Not checked
  • Clock skew beyond the tests
  • Distributed behavior outside the contract
  • Security semantics beyond existing tests
2 things need youA decisionA review
Fig. 2A pull request that says what wasn’t checked.

We start with one API that two teams share.

When an agent’s change touches an API another team depends on, the open question should go to its owners before the work runs.

When an agent-assisted change affects an API another team depends on: identify the unresolved contract decision and involve the right owner; then run, repair and verify the implementation within that decision, returning to people only for decisions and required review.

The first setup is deliberately narrow: one GitHub repository with the provider, the consumer and a versioned OpenAPI definition, and one relationship the customer nominates. Assent would run the implementation; people would approve decisions and merges. Dependencies outside the selected repository would stay explicitly outside verified coverage. None of this restricts the long-term product to API changes.

Concept diagram
One repositoryexample/commerce
RProviderReservations
CConsumerCheckout
Outside verified coverageConsumers outside this repository
Fig. 3The first setup is narrow on purpose. It doesn’t limit the product to API changes.

People decide. Assent should run the agents.

People keep the authority. Assent should do the running, and return to people for decisions and required review.

Concept diagram

People

Frame problems, resolve consequential choices and exercise authority.

accepted decisionsdecisions and required review

Assent

Is designed to prepare context, involve the right owners, turn accepted decisions into scoped agent work, run that work and repair bounded failures, verify what it can and state what it cannot, and bring people back only when a decision needs them.

Fig. 4People keep the authority. Assent should do the running, and return to people for decisions and required review.

Access to the code is not authority to change it.

Running the agents would let Assent enforce the limits its runner supports. The authority stays with the owners.

An agent can understand a change and have access to make it without having authority to change an organizational contract. Operating the agents would not remove that boundary; it would make runtime enforcement possible for the controls the runner supports.

Autonomy would be bounded by explicit grants and customer policy, and it would expand for each change class only with evidence.

Concept diagram
  1. 01ShadowFirst build
  2. 02Implementation onlyFirst build
  3. 03Policy-gated mergeLater
  4. 04Policy-gated release promotionLater

Stage labels are targets, not availability. Nothing is available yet.

Fig. 5Autonomy would grow one step at a time, per kind of change, only when you switch it on and only with evidence.

Assent has to beat your best setup, vendor tools included.

The test is total human effort per delivered change, including the time spent operating agents.

The product is worth buying only if it reduces total human effort per successfully delivered, policy-compliant change, including the effort of operating coding agents, without unacceptable delay, missed consequential decisions or unsustainable execution cost.

The comparison is your improved existing workflow, including vendor-native agent orchestration where you have it, not an unprepared agent prompt.

total human effort per delivered changeConcept diagram
  1. Your normal workflow
  2. Your improved workflowincluding vendor-native agent orchestration where you have it
  3. Assentincluding the effort of operating coding agents

What counts as effort

  • context preparation
  • exploration
  • deciding
  • approvals
  • operating agents
  • review
  • rework
  • administering the product
  • support

successfully delivered, policy-compliant change

Fig. 6Three ways, compared on the same illustrative change. No results yet.
Measured against your best setup

One record from decision to evidence.

Our bet is that linking the decision, its owner, the grant, the run and its evidence saves more effort than any part alone.

Agent vendors run agents too, and may build more of this.

Execution platforms run agents; decision tools record choices. The hypothesis is that one system of record from decision to execution, linking a decision, its authorized owner, a current scope grant, the agent work run under it and the evidence that work produced, reduces total effort beyond either alone. A decision schema, a graph or an agent runner alone is not enough.

The competition is real. Execution platforms move quickly, two of them come from GitHub and OpenAI, and many teams already pay for them. We assume no uncontested space, and vendor neutrality has to be demonstrated over time.

Expose reservation expiry to CheckoutConcept UI · illustrative data
  1. DecisionReservation expiry in the public APIDEC-0031 · rev 2
  2. OwnerRCReservations·Checkout
  3. GrantImplementationGR-7f3a
  4. RunClaude Codemanaged runner
  5. Evidenceevidence attached#218
Fig. 7A decision schema, a graph or an agent runner alone is not enough. Illustrative data.

The rules the design holds to.

People decide, and nobody is ranked. What wasn’t verified is said plainly.

Who decides

  1. Humans decide; Assent should run the agents within granted scope.
  2. Recommendations, accepted decisions and runtime grants are different things.Acceptance is not a grant
  3. Isolated experiments may inform decisions. Exploration does not authorize adoption or merge.

When people are needed

  1. Materiality depends on effects, not line count, folder name or model confidence alone.
  2. Optimize human attention: involve people for decisions, required review, and failures beyond bounded repair.
  3. Increase autonomy only with evidence, per change class and under customer policy.

What is recorded and said

  1. State what is enforced, checked and dependent on human judgment; never hide what remains unverified.Enforced, checked, or a person
  2. Preserve changes in decisions, authority and scope as explicit history.
  3. Measure teams and systems, never individual rankings.

What would prove us wrong

We will narrow or stop if any of these turns out true.

  1. No recurring problem

    Teams can’t produce recurring examples of the problem.

  2. Existing tools match it

    Improved existing tools, including vendor-native orchestration, perform as well.

  3. Buyers prefer to govern

    Buyers would rather govern their own agents than have them operated. A product that governs without running them becomes the fallback.

  4. Cost can’t be priced

    Execution cost per delivered change can’t be priced sustainably.

  5. Setup needs experts

    Setting up authority needs continuing expert consulting.

  6. No paid continuation

    Useful trials don’t lead to paid continuation.

Status of this direction

This direction was adopted on 2026-10-06, replacing an earlier governance-only thesis. No external willingness-to-pay or pilot outcome evidence exists yet. Weak demand will not be read as a request for more features; the next tests are defined in advance, with explicit continuation and stop criteria.

Adopted
2026-10-06
External evidence
None yet
Next tests
Defined in advance, with stop criteria

Test the thesis on one of your changes.

We reconstruct it with you before showing a solution: the decisions, their owners and the effort, agent operation included.