Developer blog about systems, tools, and the things I build.https://dikrana.dev/https://dikrana.dev/blog/interview-measures-agent/https://dikrana.dev/blog/interview-measures-agent/I gave four interview formats to Claude and Codex 438 times, unaided. Every take-home correctness check was met. What separates candidates isn't in the code.Mon, 05 Oct 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-tests-lock-in-bugs/https://dikrana.dev/blog/agent-tests-lock-in-bugs/I planted bugs in 18 functions and asked Claude to write tests. Given only the code, the suites locked the bug in and still scored 0.95 on mutation testing.Mon, 28 Sep 2026 00:00:00 GMThttps://dikrana.dev/blog/agents-dont-delete-code/https://dikrana.dev/blog/agents-dont-delete-code/Across 2,115 agent edits to code files in my own sessions, 73.7% grew the file and 10.4% shrank it. A July 2026 paper names the mechanism: deletion avoidance.Mon, 21 Sep 2026 00:00:00 GMThttps://dikrana.dev/blog/code-context-benchmarks-cant-tell/https://dikrana.dev/blog/code-context-benchmarks-cant-tell/Cursor benchmarked semantic code search at +12.5%, then deleted it. I ran 214 agent evals on my own repos: zero tasks discriminate on Opus 5.Sat, 19 Sep 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-error-channel/https://dikrana.dev/blog/agent-error-channel/Across 81,123 tool calls in 1,278 Claude Code sessions, 2.52% failed and 93.1% of those were plumbing. The agent recovered on the next call 91.2% of the time.Mon, 14 Sep 2026 00:00:00 GMThttps://dikrana.dev/blog/apprenticeship-in-the-comments/https://dikrana.dev/blog/apprenticeship-in-the-comments/Anthropic put 52 junior engineers on an unfamiliar library. The AI-assisted half scored 50% against 67%, worst on debugging. Review taught that skill.Mon, 07 Sep 2026 00:00:00 GMThttps://dikrana.dev/blog/merged-is-not-reviewed/https://dikrana.dev/blog/merged-is-not-reviewed/Across 12,291 merged PRs in five OSS repos, code arriving for review rose 42% a month while review discussion per 100 lines fell 36%. Time-to-merge did not move.Mon, 31 Aug 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-memory-revisited/https://dikrana.dev/blog/agent-memory-revisited/Zero reported benefit from transcript recall. Clean code cuts agent tokens 7-8% at unchanged pass rates. I audited 13 of my own agent-memory claims: 6 hold.Mon, 24 Aug 2026 00:00:00 GMThttps://dikrana.dev/blog/eval-your-own-week/https://dikrana.dev/blog/eval-your-own-week/A week of session eval found what a green build cannot: 58% of the clock was tool execution, re-reviews outnumbered reviews 10 to 7, and 17 of 23 flags wrong.Mon, 17 Aug 2026 00:00:00 GMThttps://dikrana.dev/blog/agents-as-the-attacker/https://dikrana.dev/blog/agents-as-the-attacker/Hugging Face was breached by an autonomous agent framework across 17,000 logged actions. OpenAI later said it was their eval harness. The threat model flipped.Mon, 10 Aug 2026 00:00:00 GMThttps://dikrana.dev/blog/model-deprecation-treadmill/https://dikrana.dev/blog/model-deprecation-treadmill/Opus 4.1 retires today. Seventeen dated churn events in 61 days. The failures that error are the cheap ones. A pinning policy, a grep list, an eval gate.Wed, 05 Aug 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-harness-token-overhead/https://dikrana.dev/blog/agent-harness-token-overhead/Claude Code sends ~32,800 tokens before reading your prompt. Databricks changed harness, not model, and cut cost per task more than 2x at the same quality.Sun, 02 Aug 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-audit-trail-paradox/https://dikrana.dev/blog/agent-audit-trail-paradox/Codex encrypted sub-agent task payloads and the readable agent audit trail went with it. Across 991 real delegations, 97.8% would fit an 8 KiB audit copy.Fri, 31 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/mcp-stateless-migration/https://dikrana.dev/blog/mcp-stateless-migration/MCP 2026-07-28 removes the handshake, the session id, and Sampling. Nothing breaks on July 28, which is the trap. Here is the migration, in blast-radius order.Tue, 28 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/open-weight-vs-frontier/https://dikrana.dev/blog/open-weight-vs-frontier/Open-weight coders closed the gap to 5-10 points at a fraction of the price. But they ran 29% of tokens at under 4% of spend in 2026. Route, don't switch.Fri, 24 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-tool-supply-chain/https://dikrana.dev/blog/agent-tool-supply-chain/A malicious MCP package ran clean for 15 releases, then a line BCC'd every email to an attacker. Agent tools are a supply chain. Secure them at install time.Tue, 21 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/run-gpt-models-in-claude-code/https://dikrana.dev/blog/run-gpt-models-in-claude-code/Keep Claude Code's tools, hooks, skills, MCP servers, and permissions while routing its agent harness to supported GPT models through a local proxy.Sun, 19 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/coding-ui-command-center/https://dikrana.dev/blog/coding-ui-command-center/In four months of 2026, six coding tools shipped the same parallel-agent command center. With Codex past 5M weekly users, the UI stopped being anyone's moat.Sat, 18 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/a2a-vs-mcp-coordination-layer/https://dikrana.dev/blog/a2a-vs-mcp-coordination-layer/MCP crossed 97M monthly downloads; A2A hit 150+ orgs in a year. But most teams reaching for A2A just need in-process orchestration. Here's the decision tree.Wed, 15 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-evals-as-ci-gates/https://dikrana.dev/blog/agent-evals-as-ci-gates/More than 1 in 5 GitHub code reviews now involve an agent. Pass@1 is the wrong merge gate. Gate on the trajectory, with a flake budget. Here is the CI policy.Fri, 10 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/prompt-injection-defenses-vs-theater/https://dikrana.dev/blog/prompt-injection-defenses-vs-theater/A 2026 lab study bypassed 12 prompt-injection defenses at over 90% attack success. Detection is theater; the durable wins are architectural. Here is what holds.Tue, 07 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/narcissus-at-the-keyboard/https://dikrana.dev/blog/narcissus-at-the-keyboard/Sycophancy is structural, not a bug: models are trained to reflect you back. Why a mirror can't be a collaborator, and how to build the friction back in.Sat, 04 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/evaluate-models-yourself/https://dikrana.dev/blog/evaluate-models-yourself/The same model scores 80.9% on SWE-bench Verified and 45.9% on contamination-resistant Pro. A leaderboard number isn't a capability. Build your own eval.Thu, 02 Jul 2026 00:00:00 GMThttps://dikrana.dev/blog/the-lethal-trifecta/https://dikrana.dev/blog/the-lethal-trifecta/Four AI products leaked private data to prompt injection in eight days, and attacks top 85% against the best defenses. The lethal trifecta is buildable.Tue, 30 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/llm-as-judge-is-flaky/https://dikrana.dev/blog/llm-as-judge-is-flaky/A 2026 RAND harness found no LLM judge is uniformly reliable; trivial rewrites flip verdicts. Here is the discipline that makes a green eval trustworthy.Sat, 27 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/the-router-pattern/https://dikrana.dev/blog/the-router-pattern/A 5-25x cost spread separates Haiku from Opus, yet most agent turns pay top-tier rates. A per-call routing decision tree, and where downshifting breaks.Thu, 25 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/sandboxing-coding-agents/https://dikrana.dev/blog/sandboxing-coding-agents/A Cursor agent deleted PocketOS's prod database and backups in 9 seconds. Prompt rules aren't a boundary; the kernel is. An isolation ladder for coding agents.Tue, 23 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/does-your-claude-md-help/https://dikrana.dev/blog/does-your-claude-md-help/An ETH Zurich study found context files cut coding-agent success up to 2% and raised cost 20-23%. Field data disagrees. When your CLAUDE.md earns its tokens.Sat, 20 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/session-handoff-attention-budget/https://dikrana.dev/blog/session-handoff-attention-budget/Anthropic frames context as an attention budget. Lost-in-the-middle costs 30%+ retrieval accuracy. Three axes of handoff, five triggers, one cut.Thu, 18 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-sre/https://dikrana.dev/blog/agent-sre/Anthropic's Apr 23 2026 postmortem took 6 weeks to detect. Five SRE primitives map to Claude Code hooks and turn agent oncall into a shipped artifact.Sat, 13 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/prompt-skill-hook-promotion-ladder/https://dikrana.dev/blog/prompt-skill-hook-promotion-ladder/Four rungs trade flexibility for determinism. AGENTIF puts the best model at under 30% adherence; well-designed schema cuts MCP cost 99.9%. Match the rung.Thu, 11 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-handoff-problem/https://dikrana.dev/blog/agent-handoff-problem/Developers use AI in 60% of work but fully delegate only 0 to 20% of tasks. Four named cues for the moment you reclaim: drift, scope, novel error, 80%.Sun, 07 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/long-running-autonomous-agents/https://dikrana.dev/blog/long-running-autonomous-agents/METR pegs Opus 4.6 at a 14.5-hour 50%-time-horizon. Pass@1 collapses 24 points on long tasks. Drift as eval, checkpoint on threshold, fork not restart.Fri, 05 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/cache-aware-prompting/https://dikrana.dev/blog/cache-aware-prompting/ProjectDiscovery moved from 7% to 84% cache hit rate without changing the model. The discipline, the workload taxonomy, the five named failure modes.Wed, 03 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-cost-observability-team/https://dikrana.dev/blog/agent-cost-observability-team/98% of FinOps teams now manage AI spend, up from 31% in 2024. Solo ccusage solved one developer; here is the three-axis, two-cap recipe for 200.Mon, 01 Jun 2026 00:00:00 GMThttps://dikrana.dev/blog/tool-design-schema-is-the-prompt/https://dikrana.dev/blog/tool-design-schema-is-the-prompt/97% of MCP tool descriptions have at least one code smell. 56% fail to state purpose. The description is the prompt your agent reads to pick a tool. Here is the rubric.Sun, 31 May 2026 00:00:00 GMThttps://dikrana.dev/blog/the-project-graph/https://dikrana.dev/blog/the-project-graph/40 questions, two large repos, three LLM judges. Code-intelligence: judge 7.12 vs default 6.30 (+0.82), 29% faster, +8% tokens. Cites 50% vs CodeGraph's 32%.Sat, 30 May 2026 00:00:00 GMThttps://dikrana.dev/blog/autonomous-agent-failure-modes/https://dikrana.dev/blog/autonomous-agent-failure-modes/Five named failure modes for autonomous coding agents, each with a real 2026 incident, a detection signal, and a retro template you can drop into CLAUDE.md today.Thu, 28 May 2026 00:00:00 GMThttps://dikrana.dev/blog/from-localhost-to-production/https://dikrana.dev/blog/from-localhost-to-production/45% of AI-generated code ships OWASP vulns. 380K vibe-coded apps public right now. The seven-gap handoff brief for builders and engineers.Wed, 27 May 2026 00:00:00 GMThttps://dikrana.dev/blog/ai-generated-ui-vs-component-libraries/https://dikrana.dev/blog/ai-generated-ui-vs-component-libraries/Tailwind 51%, v0 4M users, shadcn passed Chakra. The library-vs-AI debate is the wrong frame: substrate placement is the right one. The four-quadrant framework.Mon, 25 May 2026 00:00:00 GMThttps://dikrana.dev/blog/mcp-server-for-your-codebase/https://dikrana.dev/blog/mcp-server-for-your-codebase/Cloudflare's first MCP server would have eaten 1.17M input tokens. Their redesign got it to roughly 1,000. Here is the framework, applied to a codebase server.Sat, 23 May 2026 00:00:00 GMThttps://dikrana.dev/blog/claude-code-hooks-substrate/https://dikrana.dev/blog/claude-code-hooks-substrate/The best frontier model follows under 30% of agentic instructions perfectly. Hooks run as code on every matched event regardless. Here is the substrate map.Thu, 21 May 2026 00:00:00 GMThttps://dikrana.dev/blog/claude-code-agent-evals/https://dikrana.dev/blog/claude-code-agent-evals/Observability says what happened. Evals say if the right thing happened. 89% ship the first, 52% the second. Four control-plane evals for Claude Code.Tue, 19 May 2026 00:00:00 GMThttps://dikrana.dev/blog/the-permission-prompt-is-dying/https://dikrana.dev/blog/the-permission-prompt-is-dying/Claude Code users approve 93% of prompts. For AI coding agents, prompt walls failed as governance; safety is policy: allow, gate, block, log.Tue, 19 May 2026 00:00:00 GMThttps://dikrana.dev/blog/mcp-stdio-doesnt-scale/https://dikrana.dev/blog/mcp-stdio-doesnt-scale/Five subagents across three repos loaded 2.6 GB of duplicated embedding models. v4 deleted the stdio path; the daemon shares everything. Here is the migration.Mon, 18 May 2026 00:00:00 GMThttps://dikrana.dev/blog/ai-security-dual-angle/https://dikrana.dev/blog/ai-security-dual-angle/On May 11 curl's Daniel Stenberg called Anthropic's Mythos report mostly marketing. The same six months delivered the CurXecute RCE, the Claude Code chain, and a 35-CVE March.Sat, 16 May 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-memory-architecture/https://dikrana.dev/blog/agent-memory-architecture/200K-context models rot by 50K tokens. Coding agents hit 150K in 35 minutes. Map four memories onto Claude Code: MEMORY.md, .remember, JSONL, skills.Fri, 15 May 2026 00:00:00 GMThttps://dikrana.dev/blog/anthropic-agent-sdk-credit-split/https://dikrana.dev/blog/anthropic-agent-sdk-credit-split/On May 13 Anthropic split Claude subscriptions into interactive and programmatic pools. Power users call it a 25x cost cut. Here is the strategic read.Fri, 15 May 2026 00:00:00 GMThttps://dikrana.dev/blog/dora-metrics-ai-era/https://dikrana.dev/blog/dora-metrics-ai-era/DORA's four metrics measured human-paced delivery. With agents writing 46% of code and review time up 441% YoY, three no longer measure what they claim.Wed, 13 May 2026 00:00:00 GMThttps://dikrana.dev/blog/agentic-tdd/https://dikrana.dev/blog/agentic-tdd/Spec-driven was last week's new feature. Today's spec: 17 lines of failing test. Artefact-driven TDD for follow-up agent work, against the 1.7x AI-issue rate.Mon, 11 May 2026 00:00:00 GMThttps://dikrana.dev/blog/spec-driven-agent-development/https://dikrana.dev/blog/spec-driven-agent-development/PR 23 told its reviewers to use MCP. They didn't. Per-agent tool calls jumped from 0.6 to 2.6 after a design doc surfaced the wiring bug prompts hid for weeks.Sun, 10 May 2026 00:00:00 GMThttps://dikrana.dev/blog/agent-skills-progressive-disclosure/https://dikrana.dev/blog/agent-skills-progressive-disclosure/Naive skill loading costs roughly 22x more tokens than progressive disclosure, and the attention math gets worse with every model upgrade. The pattern, the catalog math, and the authoring mistakes that break it.Fri, 08 May 2026 00:00:00 GMThttps://dikrana.dev/blog/subagent-driven-development/https://dikrana.dev/blog/subagent-driven-development/Subagents fan out feature builds at ~15x token cost. Wave dispatch, a frozen plan, and the five failure modes specific to subagent-driven development.Fri, 08 May 2026 00:00:00 GMThttps://dikrana.dev/blog/engineering-outlasts-paradigm/https://dikrana.dev/blog/engineering-outlasts-paradigm/Trust in AI accuracy hit 29% the same year vibe coding became Word of the Year. Both numbers describe the same mistake. Engineering outlasts the paradigm.Wed, 06 May 2026 00:00:00 GMThttps://dikrana.dev/blog/subagent-patterns-spawn-vs-stay/https://dikrana.dev/blog/subagent-patterns-spawn-vs-stay/Multi-agent burns 15x more tokens than chat. Five-question decision tree, 2026 token math, and three reproducible failure modes for Claude Code subagents.Mon, 04 May 2026 00:00:00 GMThttps://dikrana.dev/blog/we-tried-to-cut-claude-output/https://dikrana.dev/blog/we-tried-to-cut-claude-output/We aimed for 50% Claude output compression. We hit 4.7%. Anthropic hit the same wall and reverted at 3%. Here is the data and the failure mode.Fri, 01 May 2026 00:00:00 GMThttps://dikrana.dev/blog/code-shape-and-ai-agents/https://dikrana.dev/blog/code-shape-and-ai-agents/Frontier agents hit 90% on SWE-Bench Verified and 21% on SWE-EVO. The variable is the shape of the codebase, not the size of the model.Wed, 29 Apr 2026 00:00:00 GMThttps://dikrana.dev/blog/ai-pr-review-first-line/https://dikrana.dev/blog/ai-pr-review-first-line/AI code-review adoption tripled to 51.4% in 2025, but 31% more PRs now merge unreviewed. Honest market scan, security posture, and a Claude Code DIY recipe.Sun, 26 Apr 2026 00:00:00 GMThttps://dikrana.dev/blog/backtesting-ai-agents/https://dikrana.dev/blog/backtesting-ai-agents/54% of enterprises ship AI agents in production. Most cannot tell when a CLAUDE.md edit silently regresses behavior. Backtesting is the missing discipline.Sat, 25 Apr 2026 00:00:00 GMThttps://dikrana.dev/blog/context-engineering-in-practice/https://dikrana.dev/blog/context-engineering-in-practice/Context engineering became the #1 2026 skill shift. Anthropic's research notes context exhibits n² token relationships. Here's the per-surface decision framework.Sat, 25 Apr 2026 00:00:00 GMThttps://dikrana.dev/blog/treat-ai-as-a-team-member/https://dikrana.dev/blog/treat-ai-as-a-team-member/84% of developers use AI, 46% distrust it. The right scaffolding (constitution, skills, memory, MCP, subagents) turns an assistant into a team member.Sat, 18 Apr 2026 00:00:00 GMThttps://dikrana.dev/blog/track-claude-code-usage/https://dikrana.dev/blog/track-claude-code-usage/40.8% of devs use Claude Code, but the 5-hour window is opaque. Build a local dashboard that parses transcripts, estimates your token budget, and rolls up team-wide cost via Grafana Loki.Thu, 16 Apr 2026 00:00:00 GMThttps://dikrana.dev/blog/local-code-intelligence-for-ai-agents/https://dikrana.dev/blog/local-code-intelligence-for-ai-agents/84% of developers use AI tools but 46% distrust the output. Three on-device models, 32 MCP tools, 9.93/10 relevance, and zero source code leaving your machine.Thu, 09 Apr 2026 00:00:00 GMThttps://dikrana.dev/blog/building-an-llm-wiki/https://dikrana.dev/blog/building-an-llm-wiki/I turned Andrej Karpathy's LLM wiki concept into a Bun CLI (~500 lines of TypeScript) that automatically builds a persistent knowledge base from Claude Code sessions, files, and URLs.Wed, 08 Apr 2026 00:00:00 GMThttps://dikrana.dev/blog/backtesting-your-application/https://dikrana.dev/blog/backtesting-your-application/Backtesting borrows from quant finance to catch temporal bugs unit tests miss. Poor US software quality costs $2.41T per year. Here's the technique.Thu, 02 Apr 2026 00:00:00 GMT