Agent Design Language is a deterministic cognitive architecture for building agent-based systems that are reliable, governable, observable, and reviewable.
ADL is a Rust-backed runtime and documentation system for turning agent work into explicit programs, governed tool calls, traceable artifacts, review packets, demos, and milestone evidence.
- Why ADL Exists
- What ADL Provides
- Core Ideas
- Quick Start
- Recent Demos And Proofs
- Recent Milestones
- Documentation Map
- Project Status
Agent systems are crossing from impressive prototypes into real operational infrastructure. To make that transition safely, they need more than fluent model output: they need durable programs, explicit authority, governed tools, state you can trust, and evidence strong enough for teams to build on.
As these systems take on more authority, review and security work also need to become continuous. Serious agent platforms cannot rely only on occasional audits or after-the-fact explanations; they need workflows that can find, replay, and repair weaknesses as part of normal operation.
ADL turns those requirements into an architecture for dependable agent systems:
- deterministic workflows that make agent behavior reproducible
- governed tools that separate model intent from runtime authority
- Freedom Gate policy checks before risky action
- traces, artifacts, and replay surfaces that make outcomes durable
- milestone proof packages that connect product claims to evidence
The project goal is simple: make agent-based systems safe enough to operate, clear enough to trust, and structured enough to improve.
ADL already has a substantial platform baseline:
- a Rust runtime and CLI for deterministic workflow execution
- explicit workflow, task, agent, provider, and tool artifacts
- bounded concurrency, retry, failure policy, signing, and verification
- run artifacts, traces, replay-oriented inspection, and review records
- governed tool calls through UTS + ACC
- traceable agent communication through ACIP
- Runtime v2 and CSM Observatory planning and proof surfaces
- Gödel agents and the Gödel-Hadamard-Bayes algorithm
- structured PR/control-plane workflow with SIP, STP, SPP, SRP, and SOR records
For the full capability matrix, read the canonical feature index: docs/planning/ADL_FEATURE_LIST.md.
ADL starts with a deterministic runtime. Agent behavior is represented as explicit programs, bounded state, policy decisions, and replayable artifacts so intelligence can become infrastructure instead of an unreproducible transcript.
- The ADL runtime and CSM, the Cognitive Spacetime Manifold, are the foundation: they turn agent intent into governed, replayable execution inside a persistent runtime world with durable traces, artifacts, state transitions, causality, identity continuity, and operator-visible observability.
- AEE, the Adaptive Execution Engine, is ADL's adaptation lineage: bounded strategy selection, recovery, learning, and policy-aware execution without hidden magic.
- The red/blue adversarial security model makes attack, defense, exploit replay, and purple-team coordination part of the runtime evidence story rather than a separate theater exercise.
- Gödel agents are the long-running direction for self-reference, self-improvement, and reviewable adaptation inside the deterministic runtime.
- The Gödel-Hadamard-Bayes algorithm is the cognitive loop behind that work: structured awareness, controlled hypothesis generation, and evidence-weighted judgment before authorized action.
- UTS + ACC gives the runtime governed tools: portable tool shape stays separate from permission, visibility, redaction, and audit evidence.
- ACIP gives agents a communication layer for conversation, consultation, delegation, review, handoff, and negotiation that remains traceable by the runtime.
Generate the current v0.91 cognitive-being flagship proof packet:
cargo run --manifest-path adl/Cargo.toml -- runtime-v2 cognitive-being-flagship-demo --out artifacts/quickstart/cognitive-being-flagshipInspect a current v0.91 multi-agent workflow plan:
cargo run -q --manifest-path adl/Cargo.toml --bin adl -- adl/examples/v0-91-chatgpt-gemini-claude-triad-conversation.adl.yaml --print-planThese are three high-signal recent demo entrypoints.
Start with the completed v0.91.4 C-SDLC release evidence lane:
Generate the v0.91 cognitive-being flagship proof bundle:
cargo run --manifest-path adl/Cargo.toml -- runtime-v2 cognitive-being-flagship-demo --out artifacts/v091/demo-d13-cognitive-being-flagshipRun the v0.90.5 governed-tools flagship demo:
cargo run --manifest-path adl/Cargo.toml -- demo demo-v0905-governed-tools-flagship --run --trace --out artifacts/v0905/flagship-demo --no-openRun the v0.89.1 adversarial self-attack demo:
cargo run --manifest-path adl/Cargo.toml -- demo demo-h-v0891-adversarial-self-attack --run --trace --out .adl/reports/adversarial-demo --no-openReview the v0.91 ChatGPT + Gemini + Claude triad conversation from issue #2764.
v0.91.7 is the active closeout tail feeding the required v0.91.8 bridge. WP-01 through WP-17 are closed; WP-17 closed through issue #4644 and merged PR #5539. The remaining v0.91.7 gates are WP-18, WP-19, WP-20, WP-21A, and WP-23; WP-21 and WP-22 are closed retained planning/review evidence. Closed issue state does not imply release readiness, and the milestone sprint-review register retains residual findings and non-claims. v0.92 consumes the reviewed v0.91.8 exact-revision handoff, not v0.91.7 prose as direct approval.
Start here:
- v0.91.7 README
- v0.91.7 sprint plan
- v0.91.7 issue wave
- v0.91.7 feature-doc index
- v0.91.7 sprint-review register
- v0.91.8 bridge package
- v0.91.8 to v0.92 handoff
v0.91.6 is the completed first pre-v0.92 bridge/readiness tranche. Its retained handoff and release-tail evidence are inputs to v0.91.7; they are historical records rather than the current review entrypoint.
Start here:
v0.91.5 remains the immediate upstream bridge package between the C-SDLC rollout closeout and the v0.92 first-birthday milestone. It carried multi-agent stabilization, provider/model breadth, public prompt records, demo readiness, AEE completion routing, and activation testing so v0.91.6 and v0.91.7 could open from a cleaner operational base.
Start here:
v0.91.4 is complete as the Cognitive SDLC completion and hardening milestone.
Its issue wave closed through Sprint 4 release ceremony, and the crate version
is 0.91.4. The milestone turns the v0.91.3 first slice into default
operating practice:
validator/doctor/conductor/editor alignment, signed trace and ObsMem handoff
hardening, five-minute-sprint repeatability, validation-tail/PVF work, and
bounded sidecar evidence for CodeFriend and WildClawBench.
Start here:
- v0.91.4 README
- v0.91.4 issue wave
- v0.91.4 sprint plan
- v0.91.4 demo matrix
- v0.91.4 quality gate
- v0.91.4 release evidence
v0.91.3 is complete. It proved one bounded Cognitive State Transition with tracked cards, transition evidence, review synthesis, merge-readiness truth, ObsMem handoff, and a reviewer-facing first-proof/demo package. It does not claim full C-SDLC default operation; v0.91.4 owns repeatability, enforcement, signed trace proof, validation-tail hardening, and default adoption for future software-development issues.
v0.91.2 is complete. It closed the UTS + ACC benchmark, runtime/test-cycle recovery, CodeFriend productization, Workspace bridge, modernization, publication packet, rustdoc/doc cleanup, repo-visibility follow-on, and workflow-guardrail pressure-release band. Its release-tail review, remediation, next-milestone planning, and ceremony work prepared v0.91.3 and v0.91.4.
v0.91.1 is complete. The core implementation, review, next-milestone handoff, final next-milestone review-pass wave, and release ceremony package landed before v0.91.2 opened.
It ended with an observatory-visible agent runtime proof inside the CSM boundary: lifecycle state, citizen standing/state, memory/identity architecture, Theory of Mind, capability testing, intelligence metrics, governed learning, secure ACIP/A2A hardening, and an agent-shaped runtime run that remains explicit about non-claims for birthday, identity continuity, and external federation.
Start here:
v0.91 is the completed 0.91.0 milestone. Its core line landed
the moral governance, wellbeing, kindness, humor, affect, cultivated
intelligence, structured planning, SRP, and secure intra-polis Agent Comms
band. Internal review, third-party review, accepted-finding remediation,
next-milestone handoff, release ceremony, tag, and release publication are
complete.
Start here:
v0.90.5 is the completed Governed Tools v1.0 milestone. It landed: Universal Tool Schema, ADL Capability Contract, deterministic registry and compiler surfaces, governed execution policy, trace/replay/redaction evidence, dangerous negative proofs, model compatibility work, and the first ACIP/Comms integration slice.
Start here:
v0.90.4 is the completed bounded citizen-economics and contract-market milestone. It made contract schema, bid schema, evaluation, lifecycle, transition authority, external counterparty boundaries, delegation, and one bounded contract-market proof legible without claiming payment rails or production markets.
- Feature list: canonical capability overview and roadmap truth.
- Explainers: short entrypoints for UTS + ACC, ACIP, AEE, red/blue security, Gödel Agents, and CSM.
- Docs index: repository documentation entrypoint.
- Changelog: milestone-level project history.
- ADRs: architecture decisions.
- GHB algorithm: cognitive loop behind Gödel-agent work.
- Examples: runnable ADL examples.
- Demos: demo-oriented proof surfaces.
- AGENTS.md: repository-local operating contract for coding agents.
- Active milestone: v0.91.7 closeout tail
- Required next bridge: v0.91.8 exact-revision platform acceptance and handoff
- Current ADL and Runtime v2 crate version: 0.91.7
- Independent Runtime v3 kernel package version: 0.92.0; it remains separately versioned and does not by itself claim v0.92 activation readiness
- Most recently completed milestone: v0.91.6
- Current milestone state: WP-01 through WP-17 are closed; WP-18, WP-19, WP-20, WP-21A, and WP-23 remain open; WP-21 and WP-22 are closed retained evidence
- Primary implementation language: Rust
ADL is under active development. The repository contains implemented runtime surfaces, completed milestone evidence, active milestone docs, and forward planning. Treat milestone documents as bounded engineering records: they say what is implemented, what is demoable, what is under active execution, and what remains planned.