AGENT-11

Changelog

Track the evolution of AGENT-11 with detailed release notes and feature updates.

Updated · the repository CHANGELOG is authoritative

Every entry below appears in that CHANGELOG. Two release tags in the repository have no CHANGELOG entry and so are not listed here rather than being described from guesswork; the repository tags are the complete record. There is no 3.x line — the framework went from 2.x to 4.x.

On main, unreleased

Unreleased

2026-08-03Map-First Orientation & Honest Enforcement Claims

✨Added

  • Map-first orientation across all 11 specialists and all 18 missions: Glob and Grep to locate before you read, read only the lines you need, never read a whole file to find one symbol. Orientation is the largest avoidable token cost in a session and was previously an implied habit rather than a stated instruction
  • connect-mcp and operation-recon now actually deploy. Both existed in the library but were missing from the installer’s mission list, so no deployed project ever received them. Re-run the install command to pick them up
  • Three repo checks, each silent and exit 0 when compliant: orientation coverage, enforcement claims against what the rules actually provide, and deployment coverage across both of the installer’s mission lists

🔄Changed

  • Every claim about what protects your quality gates is now scoped to what is actually enforced. The shipped permissions.deny set is four Edit() rules covering .quality-gates.json, **/*.quality-gates.json, gates/** and .gates/**. Eleven places in the library previously implied it also covered an acceptance-criteria test, a benchmark or a metric command. None of those is covered by any shipped rule. Behaviour is unchanged and the deny rules are byte-identical: what changed is that the documentation is now true
  • The Bash gate guard blocks 12 command forms, up from 5, adding rm, truncate, shred, unlink, dd of=, ln -s and in-place perl and ruby edits against gate paths. Deleting a gate passes it as effectively as lowering it
  • The guard is no longer described as closing the Bash route. It narrows it: a write through an interpreter or via a path held in a variable passes straight through, and no shell hook can catch those. Treat it as a speed bump. The enforceable guarantee is the Edit() deny rules
  • mission-optimize now creates the protection it used to assume. Phase 2 writes Edit() deny rules for the run’s own metric command, benchmark and fixtures, and proves they work by attempting an edit and attaching the refusal
  • code-review-loop defaults the critic and fixer to different models. Read-only separates the incentives; a different model separates the blind spots. A critic sharing the generator’s weights tends to return agreement rather than verification
  • Installation verification checks all 20 deployed mission files, up from 14. A failed copy of the missing six previously still reported a clean install
Minor Release

v6.2.0

2026-06-20Loop Discipline & Read-Only Verification

✨Added

  • Coordinator phase-gated meta-loop: /coord continue converges on two clean verify rounds rather than a fixed count, spends a per-phase error budget then escalates to the human instead of burning forward, and restarts from the last evidence-passed gate
  • Ratchet loops: mission-optimize isolates work in a git worktree, sets a median-of-3 baseline, makes one change on a named surface, re-measures, and keeps it only if it beats the baseline — otherwise hard-reverts
  • code-review-loop skill: a read-only critic raises evidence-backed findings, a read-write fixer addresses only those findings, repeating until two clean rounds or a cap
  • Read-only quality gates: .quality-gates.json and gates/ are unwritable by every deployed agent, enforced at the tool layer rather than by prompt convention. Scoped in the entry above: the rules cover those gate paths and nothing beyond them
  • Default-fail verification contract in tester, developer and coordinator: every success criterion starts failing and flips to pass only on captured command output
  • Bulk-ops toolkit for running AGENT-11 across multiple repositories (audit, apply-file, apply-upgrade)

🔄Changed

  • Opus tier references raised from 4.6 to 4.7 and Sonnet from 4.5 to 4.6 across the documentation
Patch Release

v6.1.1

2026-05-07Subprocess advisory cleanup

🐛Fixed

  • migrate-v5-to-v6.sh no longer prints a stale "Manual merge recommended" advisory when install.sh --upgrade invokes it as a subprocess — the warning appeared just before the success line and looked like the merge had failed
  • Standalone migrate-v5-to-v6.sh now points at install.sh --upgrade as the recommended automatic path
Minor Release

v6.1.0

2026-05-07Hardened Upgrade Path

✨Added

  • install.sh --upgrade: one command replaces the two-script v5 to v6 flow, with full rollback support
  • install.sh --dry-run to print the plan without making changes, plus --non-interactive / --batch-safe for CI
  • Surgical settings.json merge: user values win on every conflict, with backup, merge, re-validate and auto-restore on failure
  • restore-pre-upgrade.sh to undo an upgrade from backups
  • docs/UPGRADE.md and five end-to-end install fixtures

🐛Fixed

  • install.sh no longer reports "Tool deferring enabled" when it did not actually update settings.json
  • A v5 install followed by a plain install.sh no longer leaves a hybrid v6-library plus v5-residue state
Major Release

v6.0.0

2026-05-03The Lean Orchestrator

✨Added

  • Universal Router: deterministic mission routing across the mission library in greenfield, surgical and maintenance modes
  • Karpathy operating constitution: seven principles applied by every specialist, including read before writing and verify by running
  • Dynamic context loading: the coordinator reads only the files the mission mode requires
  • Phase Handoff blocks: a 5-field schema appended to agent-context.md at phase boundaries
  • Quality-gate hooks in .claude/settings.json for tsc, ruff and rubocop, plus a confirmation prompt on destructive Bash
  • Native MCP tool deferring via ENABLE_TOOL_SEARCH=auto

🔄Changed

  • library/CLAUDE.md shrunk from 575 lines to 78, and the /coord command file from 549 lines to 134
  • install.sh always installs all 11 specialists, where it previously defaulted to a 4-agent core squad

🗑️Removed

  • handoff-notes.md retired as a separate file and folded into agent-context.md
  • The .mcp-profiles/ directory and profile-switching system, replaced by native tool deferring
Major Release

v5.0.0

2025-12-31SaaS Boilerplate Killer Architecture

✨Added

  • /foundations init to create vision and PRD documents from a BOS-AI handoff
  • /bootstrap [template] to generate project-plan.md from saas-mvp, saas-full or api templates
  • /plan status and /plan phase [N] to read mission state out of project-plan.md
  • /coord continue for autonomous execution until blocked
Minor Release

v4.1.0

2025-11-28MCP Context Optimization

✨Added

  • 13 MCP profiles for context optimization
  • /mcp-switch [profile] command
Major Release

v4.0.0

2025-11-27Opus 4.5 Dynamic Model Selection

✨Added

  • Opus 4.5 for coordinator orchestration
  • Task tool model parameter (opus, sonnet, haiku)
  • Tiered model strategy documentation
Major Release

v2.0.0

2024-01-15Mission-driven orchestration

✨Added

  • Mission-driven orchestration system and the /coord command
  • Mission library with BUILD, FIX, REFACTOR and MVP missions
  • Mission template system for custom workflows
  • Multi-input document processing and mission progress tracking

🔄Changed

  • Major architecture update to support mission-based workflows
  • Documentation restructured around the field manual
Minor Release

v1.2.0

2024-01-01

✨Added

  • Advanced usage guide and multi-project workflow documentation
  • Custom squad configurations
  • Backup and restore system

🔄Changed

  • Installation system enhanced with project detection
Minor Release

v1.1.0

2023-12-15

✨Added

  • Full squad deployment: all 11 specialists
  • Individual agent profile documentation and collaboration protocols
  • Troubleshooting guide

🔄Changed

  • Improved installation reliability and project detection
Major Release

v1.0.0

2023-12-01Initial release

✨Added

  • Core squad deployment: strategist, developer, tester and operator
  • Project-local agent installation system
  • Claude Code integration, project detection and verification

Migration Guides

Need help upgrading? We've got you covered with detailed migration guides for major version changes.

v1.x to v2.x Migration Guide
Complete guide for upgrading to the new mission-based system
View Guide →
Breaking Changes in v2.1
Important changes and new /coord command syntax
View Guide →

What's Coming Next

Where v6 is heading after the bulk-migration validation pass.

Track
on GitHub

Latest plans & releases

Stay Updated

Get notified about new releases, features, and important updates.

Or follow us on GitHub for real-time updates.