Back to ACMF Maturity Model
Accelerated Training

ACMF Training & Acceleration

Agents with access to historical project context achieve autonomous operation 40-50% faster than cold-start deployments. Connect your GitHub repository to accelerate all 29 agents through the ACMF maturity levels.

How Knowledge Accelerates Training

1. Connect GitHub

Install the GitHub App on your repository

2. Ingest History

Autonoma analyzes commits, PRs, incidents, and docs

3. Knowledge Extracted

Patterns, conventions, and context become agent knowledge

4. Credits Granted

ACMF thresholds reduced — agents progress faster

All 29 Agents Benefit Automatically

Knowledge is injected at the framework level — into the RIGOR Research phase and the orchestrator's task dispatch. Every agent receives project context without any per-agent configuration. Connect your repo once, and all agents learn your codebase.

Four Knowledge Sources

Each source contributes weighted credits toward reducing ACMF progression thresholds. Incident history carries the highest weight because operational patterns are the most valuable for autonomous operation.

Git Commit History

ActiveWeight: 0.2

Codebase patterns, file hotspots, change coupling, commit conventions, directory activity, and author patterns.

File hotspots: src/api/auth.go modified in 45% of commits
Change coupling: middleware.go + routes.go change together 78% of the time
Conventions: team uses feat:/fix:/chore: prefixes in 92% of commits
BUILDMAINTAINEVOLVE

PR Review History

Coming SoonWeight: 0.15

Code quality standards, review patterns, change request hotspots, and team review preferences.

Review patterns: security-ai reviews get approved 2x faster than coder-ai
Change request hotspots: database migrations get 3x more change requests
Conventions: team requires 2 approvals for infrastructure changes
BUILDSECUREGOVERN

Incident History

Coming SoonWeight: 0.25

Operational patterns, incident frequency by component, resolution patterns, and MTTR metrics.

Incident frequency: payment service has 3x more incidents than auth service
Resolution patterns: 60% of database incidents resolved by connection pool tuning
Response time: MTTR for API incidents averages 23 minutes
OPERATEMAINTAINOPTIMIZE

Documentation Coverage

Coming SoonWeight: 0.1

Documentation presence, staleness, API docs, and architecture decision records.

Coverage: 12/20 services have README files, 5/20 have API docs
Staleness: infrastructure docs not updated in 6 months despite 200 code changes
ADRs: 3 architecture decision records found in docs/adr/
BUILDGOVERNEVOLVE

Credit System & Acceleration

Historical project data generates credits that reduce the minimum task thresholds required for ACMF progression. Credits are capped at 50% reduction to ensure agents still demonstrate competence through real task execution.

Credit Calculation

credits = (recordCount / 100) × sourceWeight × capabilityRelevance × qualityMultiplier × dateBonus

Where:
  recordCount       = number of records (commits, PRs, incidents, etc.)
  sourceWeight      = data source weight (0.10 - 0.25)
  capabilityRelevance = how relevant the source is to each capability (0.0 - 1.0)
  qualityMultiplier = data quality score (0.0 - 1.0)
  dateBonus         = recency bonus (more recent data = higher bonus)

Example: 1000 commits × 0.20 weight × 0.8 BUILD relevance × 0.9 quality × 1.2 date
       = 1000/100 × 0.20 × 0.8 × 0.9 × 1.2 = 1.73 → 1 credit for BUILD

effectiveMinTasks = max(minTasks × 0.50, minTasks - totalCredits)
                  = minimum 50% of original threshold
TransitionDefault MinTasksWith Full CreditsSpeedup
OBSERVE → LEARNING
100~5050%
LEARNING → ASSISTED
200~10050%
ASSISTED → AUTONOMOUS
500~25050%

Credits Only Reduce MinTasks

Success rate, consecutive success, max override rate, and min confidence thresholds are never reduced. Agents must still prove competence — credits only reduce how many tasks are needed to demonstrate it.

Setup Guide

Follow these steps to connect your repository, ingest your project history, and start accelerating agent training. This is the exact pipeline used in production.

Prerequisites

  • Autonoma GitHub App installed on your repository (GitHub Integration Guide)
  • Required GitHub App permissions: Contents: Read, Pull requests: Read, Issues: Read
  • Push webhook enabled (installed by default with the GitHub App)
  • Your Autonoma API token and customer ID (available in your Alpha Dashboard)
1

Trigger Initial Commit History Ingestion

After installing the GitHub App, trigger a full ingestion of your existing commit history. This processes up to 500 commits and extracts codebase patterns, file hotspots, change coupling, and conventions.

curl -X POST \
  https://www.theautonoma.io/api/v1/workflows/knowledge/ingest \
  -H "Authorization: Bearer YOUR_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "repo": "your-org/your-repo",
    "customer_id": "cust_abc123",
    "max_commits": 500,
    "force": true
  }'

Response

{
  "status": "completed",
  "repo": "your-org/your-repo",
  "commits_processed": 487,
  "knowledge_items_created": 2530,
  "credits_granted": {
    "BUILD": 1,
    "EVOLVE": 1
  },
  "duration_ms": 12450
}

Parameters:

  • repo — Full repository name (org/repo format)
  • customer_id — Your Autonoma customer ID
  • max_commits — Maximum commits to process (default: 200, max: 500)
  • force — Set true to re-ingest from scratch, ignoring the watermark (use for initial setup or re-ingestion)
2

Verify Ingestion Status

Confirm that ingestion completed and check the watermark — the SHA of the last processed commit. Future push events will only process commits after this watermark.

curl -H "Authorization: Bearer YOUR_API_TOKEN" \
  "https://www.theautonoma.io/api/v1/workflows/knowledge/status?repo=your-org/your-repo"

Response

{
  "repo": "your-org/your-repo",
  "total_commits_ingested": 487,
  "knowledge_items": 2530,
  "last_ingested_at": "2026-01-31T04:18:58Z",
  "last_ingested_sha": "6edacd828af7c1b3e9...",
  "watermark_sha": "6edacd828af7c1b3e9..."
}

Watermark: The watermark_sha tracks the last processed commit. Incremental ingestion (from push webhooks) only processes commits after this SHA. Use "force": true in Step 1 to reset the watermark and re-ingest from scratch.

3

Query Extracted Knowledge

Browse the knowledge items extracted from your repository. These are the patterns that agents receive in every task's RIGOR Research phase.

curl -H "Authorization: Bearer YOUR_API_TOKEN" \
  "https://www.theautonoma.io/api/v1/workflows/knowledge/query?repo=your-org/your-repo&limit=10"

Response (example items)

{
  "repo": "your-org/your-repo",
  "total_items": 2530,
  "items": [
    {
      "type": "file_hotspot",
      "key": "src/api/auth.go",
      "summary": "Modified in 45% of commits (218/487). Primary hotspot.",
      "confidence": 0.92
    },
    {
      "type": "change_coupling",
      "key": "src/middleware.go+src/routes.go",
      "summary": "Changed together in 78% of mutual commits. Strong coupling.",
      "confidence": 0.85
    },
    {
      "type": "commit_convention",
      "key": "conventional_commits",
      "summary": "Team uses feat:/fix:/chore: prefixes in 92% of commits.",
      "confidence": 0.94
    },
    {
      "type": "directory_activity",
      "key": "src/api/",
      "summary": "Most active directory with 340 file changes across 487 commits.",
      "confidence": 0.88
    },
    {
      "type": "author_pattern",
      "key": "alice",
      "summary": "Primary contributor (42% of commits). Focuses on src/api/ and src/auth/.",
      "confidence": 0.80
    }
  ]
}

Knowledge Types Extracted:

file_hotspot

Files modified most frequently — helps agents prioritize review and testing

change_coupling

Files that change together — agents learn which files to co-modify

commit_convention

Team commit patterns — agents follow your conventions

directory_activity

Active directories — agents understand project structure

author_pattern

Contributor patterns — agents understand team structure

4

Verify ACMF Credits Were Granted

Credits are granted automatically during ingestion. Check the ingestion response (Step 1) for thecredits_granted field. Credits reduce the minimum task count required for each ACMF level transition.

Credit Scaling by Repository Size

Repo SizeCommitsExpected CreditsMinTasks Impact
Small50-1000-1 per capabilityMinimal impact
Medium200-5001-3 per capabilityModerate reduction
Large500+3-10+ per capabilitySignificant reduction

Small repos: Repositories with fewer than ~100 commits may receive 0 credits initially. This is expected — the credit formula requires sufficient data volume. Credits accumulate as more commits are pushed. With all four data sources (commits + PR reviews + incidents + docs), even small repos reach meaningful credit levels.

5

Automatic Pipeline (After Initial Setup)

Once the initial ingestion is complete, the pipeline runs automatically. Every git push triggers the full cycle with no manual intervention required.

A

Push Event Received

GitHub webhook delivers push event to Autonoma. Contains commit SHAs and repository info.

B

Incremental Ingestion

Only new commits (after the watermark SHA) are processed. Existing knowledge is preserved and updated.

C

Knowledge Updated & Credits Granted

New patterns are extracted, knowledge items are updated, and ACMF credits are recalculated.

D

NATS Event Published

A project.knowledge.updated event is published with {repo, source, commits_processed, knowledge_items}. All agents receive the updated knowledge on their next task.

RIGOR Research Enrichment

Top 30 knowledge items (by occurrence count) injected into every agent's Research phase automatically at the framework level

Task Dispatch Enrichment

Top 20 knowledge items included in every orchestrator task payload via NATS dispatch

Verification & Troubleshooting

Verification Checklist

After completing the setup, verify each component of the pipeline is working:

Ingestion returns commits_processed > 0

If 0 commits processed, check that the GitHub App has Contents: Read permission and the repo name is correct.

Knowledge query returns items with types like file_hotspot, change_coupling

If no items returned, ingestion may still be processing. Wait and re-query.

Status endpoint shows watermark_sha matching your latest processed commit

This SHA is the boundary — only commits after it will be incrementally processed on future pushes.

Credits appear in ingestion response (for repos with 200+ commits)

Credits may be 0 for small repos — this is normal. Credits accumulate with more commits and additional data sources.

Common Issues

Re-ingestion returns 0 commits

The watermark prevents re-processing commits that have already been ingested. To re-ingest from scratch, set "force": true in the ingestion request. This resets the watermark and processes all commits again.

Credits granted: 0 for all capabilities

The credit formula divides record count by 100 and multiplies by several factors. Repos with fewer than ~100 commits may not generate enough raw value to yield credits (values below 1.0 round down to 0). This is expected — credits increase as you push more commits and when PR review and incident ingestion become available.

Push webhook not triggering incremental ingestion

Verify the GitHub App is installed with push event webhook enabled. Check your GitHub App settings under Settings → Developer settings → GitHub Apps → Permissions & events. The push event must be subscribed.

Knowledge items seem stale after new pushes

Incremental ingestion only processes commits after the watermark. If patterns shift significantly (e.g., major refactor), trigger a full re-ingestion with "force": true to recalculate all patterns from scratch.

Agent Coverage

All 29 agents receive project knowledge through two injection paths: the RIGOR Research phase (framework-level) and the orchestrator task dispatch (NATS payload). No per-agent configuration is required.

orchestrator-ai
Multi-agent coordinationGOVERN
coder-ai
Code generationBUILD
architect-ai
Architecture designBUILD
tester-ai
Test generationBUILD
review-ai
Code reviewBUILD
req-ai
Requirements analysisBUILD
planner-ai
Project planningBUILD
debug-ai
Bug diagnosisMAINTAIN
maintain-ai
Maintenance opsMAINTAIN
deploy-ai
DeploymentsOPERATE
observe-ai
MonitoringOPERATE
incident-ai
Incident responseOPERATE
backup-ai
Backup managementOPERATE
security-ai
Security auditsSECURE
cloud-waf-ai
WAF managementSECURE
govern-ai
Compliance governanceGOVERN
policy-ai
Policy enforcementGOVERN
billing-ai
Billing operationsGOVERN
optimize-ai
Performance optimizationOPTIMIZE
cost-ai
Cost optimizationOPTIMIZE
dba-ai
Database optimizationOPTIMIZE
capacity-ai
Capacity planningOPTIMIZE
evolve-ai
System evolutionEVOLVE
adapt-ai
Adaptive learningEVOLVE
training-ai
Model trainingEVOLVE
data-ai
Data pipelinesBUILD
success-ai
Customer successGOVERN
tenant-ai
Tenant isolationGOVERN
reason-ai
AGI reasoningBUILD

Best Practices for Faster Training

Do

  • Use consistent commit message conventions (feat:, fix:, chore:, etc.)
  • Ingest maximum commit history on first setup (500+ commits ideal)
  • Enable push webhooks for continuous incremental ingestion
  • Maintain documentation alongside code for higher doc coverage scores
  • Use PR reviews to build review pattern knowledge (when available)

Avoid

  • Expecting immediate AUTONOMOUS mode — credits reduce thresholds, not bypass them
  • Very small repos (<50 commits) — insufficient data for meaningful patterns
  • Inconsistent commit messages — reduces the quality score used in credit calculation
  • Disabling GitHub webhooks — agents miss incremental updates

Continue Exploring

Set up your GitHub integration to start training, or learn more about ACMF and RIGOR.