Skip to main content

Roadmap

Where AgentBoot is headed. This page is forward‑looking — themes and intent, not dated commitments; priorities shift with what teams actually need. For the full record of what has already shipped, see the CHANGELOG.

Where we are today — v0.21.0 (public Beta)​

AgentBoot v0.21.0 is a public Beta. The full pipeline works end to end today: it compiles your team's AI behavior — personas, traits, rules, gotchas, and domain layers — from one hub repo into native config for the CLI surfaces of Claude Code, OpenAI Codex, and GitHub Copilot, and delivers it to every repo as a reviewable pull request, with drift detection, managed settings, blocking compliance hooks, an MCP server, harness templates, user‑level install, and the /ab natural‑language interface. Cursor, Windsurf, Gemini, and JetBrains are supported at a community tier (native output, but advisory — not an enforced control).

Beta means the surface is real and usable, and we want it tested in the wild before we call it v1.0 GA. Expect a few rough edges — and tell us where you hit them.

Support is scoped to the CLI surface today. Broadening to the IDE and editor extensions of these tools — and to additional platforms — is on the roadmap below.

Help shape v1.0​

AgentBoot is open source (Apache‑2.0) and built in the open. During the Beta, the most useful things you can do:

  • Try it on a real repo and open an issue for anything that breaks, confuses, or surprises you — bug reports and rough‑edge reports are gold right now.
  • Tell us about your platform — especially if you live in an IDE/editor extension or a community‑tier tool. Real usage shapes what graduates to enforced support first.
  • Contribute — traits, gotchas, harness templates, and platform emitters are all welcome. Open an issue or send a pull request.

Your feedback during the Beta is what turns it into a v1.0 you can standardize on.

What's next​

Now — harden the surface and release v1.0 GA​

The near‑term goal is to take the Beta to v1.0 GA: make the shipped surface rock‑solid on real teams, close the gaps Beta feedback surfaces, and earn the "you can standardize on this" bar.

  • Beta feedback loop — triage and fix what real‑world Beta use turns up. The fastest path to 1.0 is issues from actual repos.
  • Windows support hardening — the Windows CI leg is now an enforced gate (v0.16.0); remaining live-session git‑bash/path edge cases close with real‑world Beta feedback.
  • Empirical hook verification — prove blocking‑hook enforcement (exit‑code‑2 deny) end‑to‑end on each of the three official CLIs, not just the emitted shape.
  • Import & onboarding refinement — smoother first run and a sharper import classifier so pulling an existing CLAUDE.md / AGENTS.md / rules files into a hub is close to one step.

Next — reach and depth​

  • Broader platform surfaces — official support for the IDE and editor extensions of the supported tools (beyond today's CLI surface), plus a path for community‑tier platforms (Cursor, Windsurf, Gemini, JetBrains) to graduate toward enforced support as their hook and settings surfaces allow.
  • More harness templates — additional ready‑to‑tune bundles beyond sdlc-orchestrator (e.g., API‑service, event‑processor, and data‑pipeline topologies) via add template.
  • Promotion lifecycle — proposed → incubating → stable → deprecated stages for shared artifacts, so contributions can mature in the open, with smart‑sync defaults that touch only affected repos.
  • Behavioral evaluation — a working behavioral test runner for "does the persona actually behave this way" assertions, wired into CI.
  • Deeper hook coverage — more lifecycle events per platform as each tool's hook surface grows.
  • Distributed repo operations — run a git operation (e.g., pull) across every registered repo from one command, a natural extension of hub‑and‑spoke management.

Later — org scale & governance​

  • Domain layers — fuller packaged, opinionated compliance domains (healthcare, fintech, govtech): traits + personas + gotchas + instructions, ready to adopt and tune. A generic healthcare starter pack ships today (domains/healthcare-template/ — engineering guardrails only; it does not establish HIPAA or any regulatory compliance); this item is the deeper, multi-domain build-out.
  • ADR governance — an exception lifecycle for architectural decisions, with expiry validation so stale exceptions surface instead of lingering.
  • Agent‑to‑Agent (A2A) — expose personas as A2A‑callable services, complementing MCP for multi‑agent enterprise architectures.
  • Autonomy progression — an explicit ladder from advisory → auto‑approve → autonomous, with promotion gated on telemetry and trust rather than a flag flip.
  • Org insights — aggregate, anonymized signals that help a team improve its harness (where guardrails fire, where prompts get rephrased, where cost concentrates). Designed to improve the configuration, never to score individuals — and raw prompts never leave the machine.
  • Expanded user‑scope installs — build on the v1.0 user‑level install so a solo developer can run a personal setup scoped to their OS user, without a formal org hub.
  • Additional community platforms — assess integrations for further editors and agent runtimes (e.g., OpenCode, Cline) at the community tier.

Release history​

The short version — full detail in the CHANGELOG:

  • v0.20.x — assurance hardening, honest docs, and evaluation fixes (current). Every verifier made fail-closed (telemetry signature enforcement, manifest trust posture, attestation binding, MCP digest correctness — pagination, UTF-8, secret-free spawn env); evidence-pack carries MCP provenance and states its signed state; the public docs corrected to shipped reality; a named-competitor comparison page. Then, from an assistant-driven beta evaluation run against the published docs only (v0.21.0): PR-mode preconditions asserted before the repo is touched, drift reporting that counts deletions, guardrail shadowing made a validate --strict error, provenance paths resolved against the hub, an add instruction scaffold, and a /docs/guardrails page for a capability whose authoring syntax had been documented nowhere.
  • v0.19.0 — industry-bar gap closures. MCP rug-pull defense (digest-pinned server references, mcp-pin/mcp-verify, registry provenance, pins compiled to spokes); optional in-toto/DSSE attestation next to signed sync manifests (standard predicate, SSHSIG — posture stated honestly); AGENTS.md as a first-class import input, root and nested.
  • v0.18.0 — close-out sweep. Auditor evidence-pack export (agentboot evidence-pack: one signed bundle of enforcement, drift, trust postures, guardrails, telemetry chain); AGENTS.md promoted to an officially supported first-class output (advisory enforcement class, stated plainly); a full docs/website truth-up pass; and the release process documented as a verifiable contract (docs/release-process.md).
  • v0.17.0 — tamper-evident telemetry. Org-configured central sink: hash-chained events, digest-chained + SSH-signed shipped batches, telemetry-ship/telemetry-verify, honest trust model, no default endpoint — AgentBoot still never phones home.
  • v0.16.0 — hardened assurance. An adversarial audit of our own enforcement claims, then fixes for everything it found: the Stop-hook output scan now reads the payload the platform actually sends; sibling-scope content can no longer leak into a spoke; the telemetry schema is generated from the canonical event spec; verify-manifest gained signature-strip detection and allowed-signers signer authentication with an honest trust-posture readout; the secret scan covers the full compiler input surface; SBOMs cover the full production closure; Windows is a true CI gate; publishing is decoupled from merging; and every public assurance claim is now registered against the probe that verifies it (docs/assurance-claims.md).
  • v0.15.0 — tested enforcement. The platform conformance harness executes compiled hooks (block/deny/malformed/oversized probes) per platform and emits an enforcement manifest into artifacts — the capability matrix is now a tested contract, gated in CI.
  • v0.14.0 — verifiable sync. Sync manifests carry hub provenance (commit, config + policy hashes) and a content digest; with SSH signing enabled and the signer authenticated against an allowed-signers trust root (v0.16.0), the manifest is tamper-evident — an unsigned digest detects accidental corruption only. Sync PRs carry a risk-classified change summary; verify-manifest --require-signed --allowed-signers checks it all in spoke CI.
  • v0.13.0 — org-scale import. Multi-repo import sweeps now converge boilerplate shared across repos onto one promoted org artifact carrying provenance from every contributing repo — duplicate content is never re-appended and never silently overwritten, and repo-specific residuals import normally.
  • v0.12 — enterprise hardening. The adopting‑organization batch: pluggable content scanners + blocking output scan; MCP read‑only profile + identity pinning; policy exceptions with owners and expiry; merged managed artifacts per scope; claude.settings pass‑through; import‑first sync safety; telemetry contract + telemetry-inspect; prompt‑size CI gates; SBOM + checksums on releases; the /review-ai-security persona; healthcare starter pack; enterprise operations guide; scope‑layout unification; Agent Skills + plugin spec conformance (validated in CI).
  • v0.11. The public Beta. The full compile → deliver‑by‑PR pipeline for the CLI surfaces of Claude Code, OpenAI Codex, and GitHub Copilot; cross‑platform compliance hooks; drift detection; managed settings / HARD guardrails; the /ab interface; MCP server; the sdlc-orchestrator harness template; user‑level install; security hardening.
  • v0.10. The /ab skill + MCP server; harness templates; remote‑repo import; global hub registry; smart sync.
  • v0.7–v0.9. Trait‑weight calibration; multi‑platform output (Gemini, Windsurf, JetBrains); monorepo support; cost-estimate; optimization tooling; evaluation maturity.
  • v0.5–v0.6. Cross‑platform output and richer import; enterprise governance, validation, testing, and CI.
  • v0.1–v0.4. The build pipeline (validate → compile → sync); plugin packaging; the N‑tier scope model; composition types, lexicon, and AGENTS.md.