ENTERPRISE AI ADOPTION & GOVERNANCE DOSSIER (EN / TH)
Scale AI agents across engineering and operations without losing the wheel.
Enterprise AI adoption rarely stalls because models cannot write code. It stalls at 50 to 500 seats when unscoped file mutations, 441% longer PR review queues, and unverified deployments force leadership to choose between developer velocity and operational control. Light governs every agent action through one declarative control plane that pays for itself in recovered engineering hours and avoided context waste.
Light governs the actions routed through it; native-tool restrictions ship as launch configuration, and actions taken outside Light are not covered. Harness integrations are declarative for known dialects ('integration defined in the stack' across seven harness specifications and six product families: Claude Code, Codex, Antigravity, ZCode, Muse and Pi) and no operational-history claim is made for any integration.
CHAPTER 01 / THE ADOPTION WALL
Why mid-to-large scale AI adoption hits a wall after the pilot
Giving every engineer an AI coding assistant increases individual output, but without a control plane between the agent and your repositories, the bottleneck shifts directly onto senior reviewers, release engineers, and production stability.
[UNGOVERNED STATE]
Ungoverned Agent Sprawl
Observed industry bottlenecks when agents mutate repositories and infrastructure directly
+441% PR Review Wait
Faros AI (2026) telemetry across >10,000 developers shows median PR review time jumped 441%, PRs grew 51.3% larger, 31% more PRs merged without review, and production incidents per PR increased 242.7% (X-3).
89% -> 68% Reviewed Share
A 2026 multi-year study found the share of pull requests receiving human review dropped from 89% in 2023 to 68% in 2025 as agent output outpaced senior reviewer capacity (X-4).
>40% Canceled by 2027
Gartner (June 2025) forecasts over 40% of agentic AI projects will be canceled by end of 2027 due to escalating costs, unclear business value, or inadequate risk controls (A-1).
19% Slower in Controlled RCT
METR (2025) randomized controlled trial found experienced developers took 19% longer to complete tasks with AI tools while perceiving a 20% speedup from unverified rework loops (X-10).
[GOVERNED CONTROL PLANE]
Light Declarative Governance
Every tool call, file mutation, peer check, and deployment runs inside an approved manifest
Scoped Plan Manifest First
No write tool executes until the agent declares its exact file targets, shell requests, and verification steps in a structured manifest and receives human or policy approval (L-03, L-04).
Cross-Model Peer Falsification
Before approval and before closure, Light dispatches the plan and diff to an independent model slot configured to falsify assumptions and check unimplemented promises (L-05, L-06).
Separate Ability vs. Scope
Addresses Gartner's May 2026 warning that uniform governance fails to separate an agent's ability to act from its scope of access (A-3): Light binds permissions to the active plan step.
Cryptographic Receipt Chain
Every test and build produces an authenticated receipt recording exit status and source tree byte equality at start and finish, stored in a SHA-256 hash-chained ledger (L-07, L-17).
Analyst Forecast vs. Light Architectural Position
Gartner forecasts (A-1, A-2, A-3, A-4, A-6) are independent analyst research cited with publication dates. Gartner does not endorse Light or any vendor depicted; Light's mapping to the Guardian Agent pattern (A-6) is our architectural interpretation.
CHAPTER 02 / DECLARATIVE CONTROL PLANE
One control plane across six harness families and two synchronized lifecycles
Instead of maintaining brittle per-tool wrappers or relying on developer memory, Light interposes a deterministic state machine on the MCP tool channel and binds every mutation to a live plan lifecycle.
Track A: Internal Session State Machine (L-02, 4 States)
Enforced per agent session by the Light MCP gate before any tool call is admitted
01 / boot_locked
Boot Locked
All repository reads and writes are refused until light_bootstrap declares workspace root and task tier.
02 / booted
Booted & Recalled
Read-only code intelligence and persistent KB constraints load automatically before planning.
03 / grounded
Grounded & Planned
Manifest declares exact target files, commands, and creates; peer review runs; mutations stay blocked until Go.
04 / executing
Executing (Manifest-Bound)
Only declared files and commands are admitted with step tracking; out-of-scope edits trigger immediate refusal.
Track B: Visible Work & Approval Lifecycle (L-10, 5 Stages)
Operator-visible governance record from initial proposal through cryptographic verification
STAGE 1 / proposed
Proposed
Agent submits manifest (targets, requests, creates) and design rationale for peer falsification.
STAGE 2 / approved
Approved (Go)
Operator or policy grants authority for the exact manifest; scope changes require re-approval.
STAGE 3 / executing
Executing
Step-indexed mutations, checkpoint commits, and sequenced release convoy participation.
STAGE 4 / verified
Verified
Authenticated test/build receipts (source_matches=true) plus independent verify-ship peer reconciliation.
STAGE 5 / closed
Closed
Scoped paths committed and pushed, durable KB lessons recorded, and hash-chained audit trail sealed.
Six-Layer Governance & Execution Stack (L-01 to L-22)
Every session boots into a locked gate, declares its complexity tier (patch, task, deep), and automatically retrieves project constraints, killed approaches, and past failure patterns.
Plans specify exact file targets and shell commands up front. The gate enforces per-plan write budgets, repeat-call circuit breakers, and automatic plan revocation if scope expands.
Routes design drafts and post-implementation diffs to a separate model family to surface counterexamples, missing edge cases, and unfulfilled plan promises.
Serializes multi-agent deploys through a single release convoy with post-deploy health probes, automatic rollback triggers, and dependency-ordered epic execution.
Declarative Multi-Harness Dialect Coverage (H-1 to H-7)
Integration defined in the stack across 7 harness specifications and 6 product families. No operational-history claim is made per dialect beyond Light's reference deployment.
Dual CLI and SDK specifications with shared MCP governance contract and receipt verification
[DEFINED IN STACK]
CHAPTER 03 / VELOCITY ENGINE
Governance that speeds developers up instead of slowing them down
Traditional security gates add ticket queues after code is written. Light embeds code intelligence, peer verification, and release coordination inside the agent loop so engineers ship cleaner code on the first pass.
CONTEXT COMPRESSION (L-13, L-14, E-6)
97.5% Reduction in Tool-Return Token Volume
~1,182 vs ~47,000 tokens/turn
Across 3,766 recorded turns in Light's reference deployment, tree-sitter symbol indexing and bounded slices returned ~1,182 tokens per turn compared with ~47,000 tokens for raw file reads (97.5% reduction). Qualifier (E-6): this measures what tools return to the model, not model reasoning/output tokens. On metered APIs it reduces input token spend; on flat-rate seats it prevents mid-task quota exhaustion and context-window rot.
INDEPENDENT PEER FALSIFICATION (L-05, L-06)
Automated Cross-Model Design & Ship Review
2-Stage Independent Audit
Before an operator approves a medium or large plan, an independent peer model attempts to falsify the design with concrete counterexamples and a rival architecture. Before the plan closes, a verify-ship peer inspects the actual git diff against every promise in the manifest.
RELEASE CONDUCTOR & CONVOY (L-08, L-19)
Zero Collision Deploys Across Concurrent Agents
1,400+ Sequenced Release Runs
When multiple agents finish work simultaneously, Light's release conductor queues and batches deployments through registry-defined service manifests, runs post-deploy health gates, and blocks raw ssh, scp, systemctl, or pm2 improvisation.
INSTITUTIONAL MEMORY (L-15, L-16)
Killed Approaches Never Repeat Twice
Auto-Injected on Smart Index
Every closed plan records architectural constraints, root-cause diagnoses, and killed approaches into the project knowledge base. Subsequent sessions receive matching constraints automatically when navigating those files.
CHAPTER 04 / SELF-FUNDING ECONOMICS
Why Light practically pays for itself: the four cost lines and live workload estimator
Light acts directly on the four largest hidden cost centers of scaled AI engineering (E-1). Enter your organization's own seat count, compensation, billing model, and target subscription budget below to calculate your break-even threshold and recoverable engineering capacity.
Transparency & Claim Policy (E-1, E-6): The four cost lines below are structural mechanisms supported by reference-deployment telemetry (L-14) and cited industry benchmarks (X-3, X-4, X-6). All dollar outputs in the interactive calculator are computed solely from the workload assumptions you enter below, never presented as fixed vendor pricing or guaranteed savings.
Raw file dumps flood context windows. Light's 97.5% reduction in tool-return tokens (L-14) cuts metered API input spend directly; on flat-rate seats (E-6), cash token savings are $0, and the gain appears as roughly 3x-5x more active engineering tasks before hitting seat rate limits.
Industry PR review times jumped +441% (X-3) and unreviewed merges rose as human review fell from 89% to 68% (X-4). Manifest-bound scope, pre-build peer falsification, and verify-ship checks shrink diff bloat before a human reviewer opens the PR.
COST LINE 03 / UNSCOPED CHANGE INCIDENTS (E-1, X-3, X-6)
Production Rework & Outage Exposure from Out-of-Scope Mutations
Incidents per PR rose +242.7% with ungoverned AI adoption (X-3), while enterprise downtime exceeds $300,000/hour for 91% of mid/large organizations (ITIC 2024, X-6). Light refused 8,600+ out-of-scope tool calls before execution in its reference deployment (U-3).
Replaces ad-hoc deployment scripts and manual compliance evidence gathering with registry-driven release convoys, automated health verification, and hash-chained execution receipts.
Interactive Buyer Workload & Break-Even Estimator
Adjust any parameter to model your engineering and SRE organization. Keyboard arrow keys work normally inside all inputs.
Agent Seat Billing Mode (E-6)
[READY · BUYER-ENTERED ASSUMPTIONS]
Break-Even Time Saved / Seat / Month
30 minutes / month
At your entered hourly cost and target seat budget, Light pays for a seat the moment it saves this many minutes of engineering time per month.
Estimated Gross Monthly Value Recovered
$38,158
Estimated Net Annual Impact (After Budget)
$427,896
Estimated Value-to-Budget Multiple
15.3x
Recovered Engineering Hours / Seat / Month
6.6 hrs
Line 1: Metered Tool-Return Token Savings (E-6)
$0
Line 2: Senior Review & Rework Capacity Recovered
$30,310
Line 3: Avoided Unscoped-Change Incident Cost
$5,250
Line 4: Release Convoy & SRE Triage Capacity
$2,598
Flat-rate mode active (E-6): Line 1 cash savings are set to $0 because seats pay a fixed monthly fee. The 97.5% tool-return reduction instead preserves seat quota headroom and avoids mid-session lockouts.
CHAPTER 05 / PROOF, MARKET COMPARISON & CISO CONTROL MAP
Reference-deployment telemetry, 6-category market landscape, and CISO control mapping
Every capability claim is anchored to an inspectable mechanism in the Light stack, reference-deployment telemetry since March 2026, or documented public sources.
Attribution Rule (B-10, B-11): Telemetry below comes from Light's own production reference deployment running since March 2026 across five harness families. It is not presented as third-party customer telemetry. "Plans closed" includes aborted plans, and "tool calls refused" are out-of-scope policy refusals prior to execution, not blocked cyberattacks.
Six-Category Market Comparison (As of October 2026)
Filter by capability status to inspect how Light compares with single-vendor hooks, LLM proxy gateways, code-review bots, and workflow frameworks.
No market categories match the active filter. Reset the filter to inspect all 6 categories.
Market Category & Representative Tools
Pre-Execution Tool Gate
Manifest-Bound File Scope
Cross-Model Peer Falsification
Source-Verified Receipts & Release Convoy
Light Declarative Governance Control Plane (6 harness families)
[PARTIAL] Programmable graph nodes and human-in-the-loop breakpoints for custom apps
[NOT ESTABLISHED] SDK for building bespoke agents; does not govern off-the-shelf coding CLIs
[PARTIAL] Custom multi-agent graphs possible if engineered and maintained by customer
[NOT ESTABLISHED] No turnkey engineering CLI governance, smart-index, or release conductor
Showing all 6 documented market categories from 05-competitive-comparison.md (as of October 2026). Status labels follow strict claim-register definitions: [DOCUMENTED], [PARTIAL], [NOT ESTABLISHED].
CISO Capability-to-Control & Regulatory Mapping (C-1 to C-11, R-1 to R-4)
Scope Note: This table maps technical capabilities in Light to enterprise security controls and regulatory articles (EU AI Act, NIST AI RMF, ISO/IEC 42001, SOC 2 / DORA / SR 11-7). It is an architectural capability-to-control mapping for security review, not a legal compliance certification.
Control ID & Threat Vector
CISO / Regulatory Requirement
Light Enforcement Mechanism
Aligned Framework Reference
C-1 / Unbounded Action Scope
Separate agent ability to act from approved scope of access (Gartner A-3)
Manifest-bound file targets[], command requests[], and creates[] enforced per plan step (L-03, L-04)
NIST AI RMF Govern 1.2 / Map 3.5 (R-2)
C-2 / Unapproved Production Mutation
Mandatory human-in-the-loop authorization before high-impact changes
Session boot_locked/grounded gate blocks all writes until operator Go or WebAuthn hardware approval (L-02, L-10, L-22)
EU AI Act Art. 14 Human Oversight (R-1)
C-3 / Tamperable Execution Logs
Non-repudiable audit trail of every agent action and command result
Bound blast radius of stuck or looping autonomous sessions
Per-plan patch budgets, repeat-call circuit breakers, and automatic plan revocation on scope drift (L-11, L-12, L-20)
NIST AI RMF Manage 2.4 / ISO 42001 A.8 (R-2, R-3)
Executive & Engineering Persona Value Matrix
CTO / VP Engineering
Scale AI velocity across 50-500 engineers without PR gridlock
Converts chaotic agent output into manifest-scoped, peer-falsified changes that senior engineers can review in minutes instead of hours.
CISO / Head of Security
Pre-execution policy enforcement and tamper-evident auditability
Replaces advisory markdown rules with a hard pre-execution gate, RBAC/WebAuthn approvals, and SHA-256 hash-chained receipts mapped to EU AI Act Art. 12/14 and NIST AI RMF.
CFO / FinOps Lead
Predictable AI unit economics and measurable payback
Cuts tool-return context volume by 97.5% (L-14) and recovers senior review and incident rework hours so the control plane pays for itself within minutes saved per seat.
Platform / SRE Lead
Eliminate wild-west SSH/Docker commands and deploy collisions
Routes all deployments through declarative service manifests and a sequenced release convoy with automated health checks.
Staff / Senior Developer
Fast codebase navigation and zero repeat debugging
Sub-second tree-sitter symbol navigation (light_smart_index) with automatic recall of past architectural constraints and killed approaches.
CHAPTER 06 / 30-DAY SHADOW-TO-ENFORCE ROLLOUT
Adopt Light in 30 days without disrupting active sprints
Start in read-only observation mode on one engineering squad, measure your baseline token and review metrics, then enable manifest gates and release orchestration step by step.
WEEK 01 / SHADOW & BASELINE
Connect Harnesses & Baseline Telemetry
Attach existing Claude Code, Codex, or Antigravity seats to the Light MCP bridge. Enable tree-sitter smart index (L-13) and measure baseline tool-return token volume and PR review cycle times.
WEEK 02 / MANIFEST & PLAN GATE
Enable Scoped Plans & Write Admission
Turn on session boot lock and manifest-bound planning (L-02, L-03) for medium and large tasks. Out-of-scope file mutations and raw destructive shell commands are refused before execution.
WEEK 03 / PEER FALSIFICATION & RELEASE CONVOY
Wire Cross-Model Peer Review & Service Registry
Configure a second model family for pre-Go design falsification and verify-ship checks (L-05, L-06). Register deploy targets in the release conductor (L-08).
WEEK 04 / EXECUTIVE AUDIT & ROI READOUT
Compare Measured Velocity, Refusals & Cost Lines
Export hash-chained execution receipts, refused-call audit logs, and measured review/context savings across the four cost lines (E-1) for CTO, CISO, and CFO sign-off.
Ready to evaluate Light on your engineering organization?
Walk through the architecture, inspect live reference-deployment telemetry, or scope a 30-day shadow-to-enforce pilot with Factor IO.