Found my old CV to be kinda mehh So made it into an app 😎 i-technology.net/

Latest update to the desktop and mobile version now integrates a custom image viewer for past projects with nice scroll, lightbox, and some other stuff (GPT 5.5 xHigh)

Having fun with codex, working on my agentic harness ! The new workflow is functioning well now, and the harness is moving along at it's snail pace ^^ (hundreds of iterations in, and enterprise level with auditability & accoutability) To save context and let the agent work better, i made a internal helper tool that includes AST, searching frontmatter, and other things, instead of the Agent doing a bunch of grep all over the place This allows the agent to quickly look up impacted code and stuff when doing modifications/refactors The tool is 100% cli & Agent oriented with structured JSON for everything (not made for humans) And now the tool just got a support ticket handler, so whenever the agent thinks the tool sucks, it can just open a ticket, and i can later work on that, or delegate it 🤣

And this is the current functionality matrix of our harness (already implemented) It's very "enterprisy level" oriented ..thus taking quite a while to get right Accountable work records: Runs produce reviewable records of what happened: goals, decisions, tools, policies, models, outcomes, and evidence pointers. No-secret evidence: Proof is retained without storing raw secrets, provider payloads, transcripts, raw tool arguments, or sensitive command output. Claim-to-proof closeout: Important claims must point to concrete proof: tests, live checks, audits, redacted evidence, or explicit deferrals. Policy decision history: Allow, deny, redact, skip, approval-requested, and failure states are recorded as first-class facts. Action review gates: Risky actions can be routed through explicit review instead of being silently executed. Pending action resolution: Paused actions can be approved, denied, expired, or cancelled, with single-use execution of the exact reviewed action. Fail-closed behavior: Malformed, unavailable, unsafe, timed-out, mismatched, or unauthorized paths stop safely instead of guessing forward. Access enforcement: Read/write/command boundaries are enforced at runtime, not treated as advisory labels. Work mode controls: Modes like planning, asking first, read-only, and higher-trust execution are separate from tool approval and access scope. Checkpoint and resume: Work can pause at a known checkpoint and resume through validated authority instead of restarting blindly. Tool-requested pauses: Tools can explicitly ask the work loop to pause when human input, policy, or external conditions require it. Decision requests: Non-tool decisions can pause work, wait for input, apply declared defaults, timeout, cancel, or resume with evidence. Delegated decision handling: A bounded delegate can answer scoped decision prompts without gaining tool approval, access, or human authority. Reviewer and delegate profiles: Reviewer and delegate roles are explicit slots with readiness, routing, fallback, and failure evidence. Model routing evidence: The requested model, selected model, provider route, fallback behavior, and readiness state can be recorded without exposing sensitive payloads. Work-record readback: Operators can inspect concise, no-secret summaries of what happened and why. Stop reasons: Runs distinguish done, blocked, interrupted, budget exhausted, policy pause, decision pause, and tool pause. Budgeted continuation: Longer work is moving toward explicit continuation budgets instead of open-ended autonomy. Autopilot with review boundaries: Near-term autonomous continuation composes with action review and policy gates instead of bypassing them. Delegated autopilot boundaries: Delegation can help with bounded decisions, but high-risk actions and human-required gates still block. Surface-control readback: Operator-facing controls will show effective work mode, access, review, decision, and delegate state. Audit matrices: Important control axes are checked against runtime behavior, tests, evidence routes, and named deferrals. Traceable evidence routes: Current status, decisions, closeouts, and proof artifacts link forward and backward so context can be reconstructed later.

Asked GPT 5.5 Pro Extended what it thinks of my current workflow (gave it full docs, mermaid diagrams etc) Looking good 😎 (now back to work !) --- question --- this is the current workflow we have adapted for our dark-horse project lets verify what online reasearch papers on workflows, agility, and agentic coding say, and see how our workflow fits in with these to see what we are doing well, and what we could be doing better the idea is human/agentic collaboration in an agile manner, that lets us pretty much move any product forward with as little friction as possible the idea is to later replace decision trees & various research related elements with specialized agents, that can step in when the human operator is absent, or chooses to delegate tasks but for now it remains human driven lets check latest information on agentic harnesses, systems, and how research fits in with what we are currently doing, and see what can be enhanced provide multiple paths/options each with their pros/cons

Sorry, incomplete ..but hey, i'm sharing ^^ (culmination of almost 8 months of trial/error) Now i just need to finalize it and package it in a self-contained skill TL: Lifecycle Walkthrough TR: Decision Trees BL: Reviewed Sprint Approval And Closeout BR: Adaptive Audit And Wave Transition

Uncle Bob works well and has hundreds of sprints under it's belt, but was too rigid This new version makes everything dynamic and variable, with better test/research/review gates