Research and engineering publication

AI SDLC: research, methods, and engineering practice.

LockedIn Labs’ knowledge publication for agentic engineering: original analysis, AI SDLC methods, specifications and practical reference tools.

Inspect the decision. Change the boundary.

A synthetic, browser-local demonstration of the existing Control Plane Lab engine. It sends no tool calls and grants no permission. Verifier and evidence states are declared fixture inputs.

One synthetic task; four declared conditions
ConditionChanged inputReference outcome
Within scopeDeclared prerequisites satisfiedSimulated allow
Identity revokedIdentity standing → revokedBlock
Evidence missingEvidence package → missingHold
Policy prohibitsPolicy disposition → prohibitedBlock
  1. Declare intent — Inspect the use case, accountable owner, risk inputs and exact source identity. Plan + system passport.
  2. Challenge the boundary — Change declared conditions and inspect the deterministic decision and obligations. Unsigned simulation trace.
  3. Prepare implementation — Translate the dossier into control adapters, operating duties and evidence contracts. Implementation references.
  4. Inspect executed evidence — Read the historical Java defect, failed test and passing repair, including the proof limits. Recorded engineering example.

Read the presentation guide

LockedIn Labs / Knowledge & reference

AI SDLC within
agentic engineering.

AI SDLC is an integrated capability of the Agentic Engineering Platform by LockedIn Labs (working name). AISDLC.ai is the knowledge and reference publication, with methods, specifications and tools for teams to inspect and apply.

The AI software development lifecycle connects intent, architecture, implementation, independent verification and operations. For agentic work, that includes the identities, context and tools through which agents act, and the feedback used to improve the system.

Read the AI SDLC guide

The enterprise journey: governed autonomy

These are four proposed operating modes, not a certification or a timetable. Choose by workflow and consequence; a mature hybrid model can be the right destination. An enterprise may use several modes at once.

Start with the delivery system you have.

Conventional product development is the baseline: teams define intent, build, test and operate software. Measure its outcomes and failure modes before changing who can act or decide.

Hybrid AI SDLC

AI contributes analysis, implementation and review inside a delivery process whose consequential decisions remain with people. Model judgment is advice; evidence and accountable approval decide promotion.

The human role
People set acceptance criteria, inspect the evidence and authorize consequential changes. Start with bounded work that exposes where the model helps and where it fails.
Delegated authority
Draft code, tests, documentation and recommendations within declared data and tool boundaries. Production changes require the established human approval path.
The learning boundary
Teams curate recurring failures into context and evaluation cases. Changes to the operating process remain reviewed and versioned.

Evidence before widening scope

  • A delivery baseline for lead time, escaped defects, rework and operating cost.
  • Versioned acceptance criteria and checks that can challenge the model’s work.
  • Recorded inputs, actions and outcomes with an accountable owner for each workflow.
Build a workflow adoption plan

Supervised agentic delivery

Agents complete bounded tasks across tools, rather than only suggesting the next step. A person reviews the consequential promotion, supported by an inspectable record of execution and independent checks.

The human role
An operator supervises the task and approves consequential promotion. Review the candidate and its evidence, with a tested route to reject, interrupt or recover the work.
Delegated authority
Plan, implement, run permitted checks and assemble evidence within a task contract. The agent may prepare a release candidate; it cannot grant itself release authority.
The learning boundary
Execution records can generate improvement proposals. A separate review and evaluation path decides which proposals enter the next version.

Evidence before widening scope

  • Repeated trials that include difficult cases, tool failures and interrupted execution.
  • Checks tied to the exact candidate, with protected acceptance criteria and provenance.
  • An exercised stop, escalation and recovery path, with observable operator workload.
Inspect the implementation contracts

Delegated autonomy

Preauthorized, low-risk workflow classes execute without an approval for every action. Protected checks enforce the boundary while an accountable operator monitors outcomes, exceptions and recovery.

The human role
People authorize the operating envelope, monitor exceptions and outcomes, audit a risk-based sample and retain the power to revoke authority. Their attention goes to changing conditions and consequential exceptions.
Delegated authority
Complete approved workflow classes through independently enforced checks, permissions and resource limits. Pause when evidence is missing, policy is exceeded or a consequential judgment falls outside the contract.
The learning boundary
Feedback can improve a candidate workflow within its existing scope. The executing agent cannot widen its own authority or change protected policy; scope expansion needs a separate authorization decision.

Evidence before widening scope

  • Reliable outcomes across representative trials, with false approvals and missed failures measured.
  • Decision records linking policy, model and context versions to evidence, action and observed outcome.
  • Exception thresholds, an accountable response owner and rehearsed revocation and rollback.
Study the verification and release contract

Adaptive AI factory — optional destination

An optional destination for eligible workflows: the delivery system uses outcome evidence to propose improvements to its context, orchestration and evaluations, then tests and promotes those changes through a governed path.

The human role
Accountable owners set objectives, limits and change policy. They oversee aggregate outcomes, drift and exceptions while an independent control path governs improvement and deployment.
Delegated authority
Run eligible work and propose delivery-system changes inside a defined envelope. Independently evaluated candidates can progress through preauthorized rollout checks; the agent cannot widen its own authority or change protected policy.
The learning boundary
Learning produces a versioned candidate, not an automatic rewrite of production behavior. Context updates, orchestration changes and model retraining are distinct change types with their own acceptance evidence.

Evidence before widening scope

  • An independently maintained evaluation set, with held-out cases and regression protection.
  • Versioned improvement candidates, comparative trials and monitored rollout with rollback.
  • A protected governance boundary and ongoing audits of outcomes, drift and delegated scope.
Map the agent estate and its authority

Read the AI SDLC transformation workbook

The reading room

A method for the next era of software

LockedIn Labs proposes three engineering positions informed by primary sources. Establish their value through evaluation in your own system.

Context is infrastructure.

Treat repository knowledge, tool contracts and system state as a maintained interface. Give the agent a small, traceable working set, then refresh it when the assumptions behind the task change.

Put it into practice

  1. Version task contracts, domain constraints and tool schemas alongside the system.
  2. Retrieve relevant evidence on demand; record its revision, origin and access scope.
  3. Invalidate stale assumptions and test whether a task can recover after a context reset.
The tradeoff
Curation adds retrieval latency and maintenance work. More context can increase exposure and bury useful signal.
What to measure
Track stale-source incidents, retrieval relevance and task recovery after a context reset.

Primary-source reading: Effective context engineering for AI agents (Anthropic / 2025)

Verification is a separate plane.

Design the acceptance contract before the implementation. Separate the path that generates a candidate from the path that challenges it, and judge observable outcomes rather than persuasive narration.

Put it into practice

  1. Define observable acceptance criteria, regression cases and known failure modes.
  2. Run checks in isolated trials; calibrate judgment-based grades against human examples.
  3. Link findings to the candidate artifact; name the owner of release policy and exceptions.
The tradeoff
Rigid checks can reject valid solutions. Model-based judgments add flexibility, but require calibration and remain fallible.
What to measure
Track repeated-trial reliability, escaped defects and disagreement between automated and human grades.

Primary-source reading: Demystifying evals for AI agents (Anthropic / 2026)

Autonomy has a budget.

Delegate an outcome within an explicit operating envelope: permitted actions, resource limits, stopping conditions and a recovery path. Expand authority only when task evidence supports the decision.

Put it into practice

  1. Declare the allowed tools, data boundary, spend limit and accountable owner.
  2. Enforce limits outside the model; stop or escalate when the envelope is exceeded.
  3. Record consequential actions and exercise the recovery path before widening authority.
The tradeoff
Narrow authority limits the tasks an agent can finish. Broader authority increases the consequences of mistakes.
What to measure
Track verified outcomes per unit cost, boundary violations and time to restore a safe state.

Primary-source reading: Intelligent AI Delegation (Tomašev, Franklin & Osindero / 2026)