Skip to content
Log in Request access

Lifecycle governance for agent-managed development

Ship with agents. Keep the reins.

Project tools were built for people who remember context, report their own progress and know when to stop. Coding agents do none of that. LGAM holds them to scope, routes every consequential choice to a human, and refuses “done” without evidence.

Request access See what changes

LGAM, said “l-jam”, is لجام: Darija for the bridle that steers a horse without taking its power away.

What changes

Your tracker assumes a person is doing the work.

Agents are fast, tireless and sure of themselves. They also forget everything between sessions and will report success they never checked. Six assumptions stop holding.

  1. TraditionalThe assignee remembers why.

    LGAM: Context is a contract, read before every write.

    A new agent session starts with no memory of the last one. Every LGAM project carries a live agent contract: absolute rules, guardrails and mandatory references that agents fetch before they touch code.

    Agent contract · v3
    rule
    Never settle a decision as an agent
    rule
    Do not mutate immutable process records
    guardrail
    Ground work in project knowledge before writing
    reference
    Approved client requirements

    lgam_get_agents_md

  2. TraditionalStatus is whatever someone reports.

    LGAM: State is what the server evaluates.

    Lifecycle gates run on every transition against the ledger as it stands. The same state always gives the same answer, however confident the request sounds.

    executing → review
    no-open-blocking-ods
    pass
    child-tasks-done
    pass
    tasks-have-evidence
    fail
    agent-review-approved
    pass

    409 FEATURE_GATE_FAILED

  3. TraditionalDone is a box someone ticks.

    LGAM: Done is a claim that needs evidence.

    A finished task with no linked intervention blocks its feature. Agents report what they verified and, just as plainly, what they did not.

    DEMO-12 · Export endpoint
    status
    done
    evidence
    INT-48
    verified
    41 unit tests pass
    not verified
    Letter page size

  4. TraditionalDecisions happen in meetings and threads.

    LGAM: Decisions are records with a named owner.

    An Open Decision blocks the work it governs until its authority settles it. Agents can open and analyse one. Only a human answers, and settled choices become ADRs.

    OD-3 · Who may export the full journal?
    state
    open
    blocks
    DEMO-7
    authority
    Product owner
    agent settles
    refused

    403 DECISION_AUTHORITY_REQUIRED

  5. TraditionalHistory is whatever the ticket says today.

    LGAM: History is append-only.

    Interventions, decisions and ADRs are never edited. A correction is a new record that supersedes the old one, so the trail shows what was believed and when.

    Delivery journal
    INT-46
    Page breaks verified
    INT-47
    Supersedes INT-46: verified on A4 only
    INT-48
    Export endpoint shipped

    corrections supersede · nothing is rewritten

  6. TraditionalThe tool is a board for people to click.

    LGAM: Agents are first-class users, inside a fence.

    Coding agents reach the same ledger the team reads through 84 MCP tools in five packs, with scoped keys and explicit project grants. The tools come with the rules attached.

    MCP key · tool packs
    core
    on by default
    planning
    opt-in
    governance
    opt-in
    projects
    explicit grants only

    84 tools · 5 packs

The name

Two reins. One delivery record.

Control without memory is blind steering. Memory without control is a runaway with a diary.

Govern comes from the Latin gubernare: to steer.

Read the name twice. Both readings are official.

Lifecycle Governance for Agent- Managed development

Control plane

Govern the lifecycle.

  • Five-level work breakdown with dependencies
  • Feature lifecycle with server-evaluated gates
  • Open Decisions that block what they govern
  • Requirement alignment and convergence checks
  • Verified interventions and releases

Knowledge plane

Ground every claim.

  • Versioned client requirements
  • Immutable decisions and architecture records
  • An append-only delivery journal
  • Specifications, wiki and runbooks with governed attachments
  • Hybrid search and a knowledge graph that cites its sources

The lifecycle

Hard edges for agents. Authority for humans.

Every feature moves from shaping to executing to review to done, and the server evaluates each move. Pick who you are, then try to ship it.

DEMO-7

Export the delivery journal to PDF

  1. shapingHuman accepts the shape
  2. executingAgent hands off
  3. reviewHuman accepts the result
  4. done

Gate · shaping → executing

human only
  • no-child-tasksrecordedFeature has 2 child task(s)
  • no-open-blocking-odsrecordedBlocked by OD-3

You are an agent holding an API key. Try to move the feature forward, then switch roles and try again.

What the ledger holds

  • OD-3 · Who may export the full journal? human authority
  • Child tasks · Render journal pages, Export endpoint
  • Evidence · intervention INT-48
  • Aggregate review · Spec and Standards
  • Release record · intervention INT-49

The simulator runs the transition table and checks of the LGAM API, with its error codes and messages.

The ledger

The session ends. The ledger doesn’t.

Agents read the contract, search the record and append what they did: the outcome, what was verified, what was not, and the trap the next session should avoid.

agent session · DEMO-7
  1. lgam_get_agents_mdcontract v3 · 3 absolute rules · 5 mandatory references
  2. lgam_search_project_entities “journal export”4 matches · OD-3 is open and blocks DEMO-7
  3. lgam_get_feature_context DEMO-7executing · 3 gates failing · playbook attached
  4. edits 6 files · 41 unit tests pass
  5. lgam_link_task_evidence DEMO-11, DEMO-122 tasks → INT-48
  6. lgam_submit_interventionINT-48 appended to the journal

Delivery journal

INT-48appended

Journal export endpoint and PDF renderer

Verified
41 unit tests pass; a 300-entry journal exports as one PDF.
Not verified
Letter page size. Only A4 was rendered.
Trap
The renderer streams. Buffering a whole journal exhausts memory on large projects.
INT-47supersedes INT-46

Correction: page breaks verified on A4 only

INT-46superseded by INT-47

PDF page breaks verified

The engineering plane of AI governance

Governance that can say no.

Enterprise AI governance architectures describe a chain from policy to evidence. LGAM implements that chain for the AI system most frameworks forget: the coding agent that builds the software. Each step is a ledger record, and the server evaluates the gate between them.

  1. Policy & requirements

    Strategy, policy, legal and regulatory needs

    A live agent contract per project, plus versioned client requirements. Agents read both before any write.

    • agent contract
    • client requirement

  2. Controls & procedures

    Standards, guardrails, approvals

    Lifecycle gates the server evaluates on every transition. The same state always gives the same answer.

    • no-open-blocking-ods
    • child-tasks-done
    • tasks-have-evidence
    • agent-review-approved

  3. AI system

    The tool that changes the code

    The coding agent, scoped by a short-lived key, explicit project grants and tool packs. An ungoverned agent session is Shadow AI inside the delivery pipeline. This is the sanctioned lane.

    • API key
    • project grant
    • tool pack

  4. Decisions & actions

    Human review, actions taken

    Open Decisions and ADRs. Agents open and analyse; only a human settles. ADRs are append-only and carry the trap: the mistake the decision exists to prevent.

    • OD
    • ADR
    • human acceptance

  5. Logs & evidence

    Record activity, detect issues

    Interventions record the outcome, what was verified and what was not. Evidence links to the task. The journal does not rewrite.

    • intervention
    • evidence link
    • journal

  6. Audit & assurance

    Independent review, validation

    The integrity report and alignment runs check the ledger against requirements and surface tasks that claim done without evidence.

    • integrity report
    • alignment run

  7. Findings & remediation

    Identify gaps, take corrective action

    Remediation reopens execution with a note and invalidates the review receipt before any code is repaired. Corrections supersede; they never edit.

    • review → executing
    • supersedes

NIST AI RMF

Govern
agent contract, human acceptances
Map
work breakdown, Open Decisions
Measure
alignment runs, integrity report
Manage
remediation, superseding records

ISO/IEC 42001

Management system
one ledger per project, with a live contract
Operation
lifecycle gates and evidence links
Improvement
journal and ADR amendments

EU AI Act

Record-keeping
interventions and the journal
Human oversight
the two human acceptances
Risk management
Open Decisions and blocking gates

These are mappings from ledger records to published obligations, not claims of conformity or compliance.

Governance without ceremony.

Strict where ambiguity creates rework. Quiet everywhere else.

  1. Human decisions stay human.

    Agents can surface and structure a choice. Named authority answers it and owns the consequence.

  2. History does not rewrite.

    Decisions and interventions are append-only records. Corrections supersede; they do not erase.

  3. Current state stays explicit.

    The ledger carries today’s implementation facts, open work and known deviations, so nobody reconstructs them from chat.

Put a lifecycle around your next build.

Tell us what you are building, how your agents work, and where governance breaks down today.

Request access Log in

Opens a pre-addressed access request.

LGAM · AI governance for agent-managed software delivery