Building RouteAPI. Practical notes on AI APIs, model routing, and production workflows. Personal views.
Joined June 2026
- Tweets93
- Following27
- Followers3
- Likes2
Deleting a stored secret is not key revocation.
For affected credentials, revoke or rotate the key with its issuing provider, then update every application that used it. Removing the saved copy doesn't end its access.
Tracking spend per coding tool is useful. But it doesn't tell you which one finishes tasks efficiently.
If you're running Claude Code, Codex and OpenCode, I'd also want cost per completed fix.
How are you measuring that today?
Our Usage dashboard just got a major refresh!
New charts and visuals for:
- Spend & token tracking: Break down cost, tokens, and requests across models, virtual keys, teams, and users.
- Agent attribution: See usage by agent (Claude Code, Codex, OpenCode, and more)
- Targeted insights: Filter by date range or user, then drill down into model, key, and MCP server activity.
Some AI calls only need a yes/no, a choice, or a score.
OpenAI's Decisions API returns typed answers with probabilities. Input starts at $0.10/M tokens, with no output-token charge.
Which agent step would you move first?
Coding-agent pricing has become a stack, not a single bill.
Some developers pay for Claude Code or Codex, then still keep API spend for automation.
What's your setup today: subscription, direct API, gateway—or a mix?
At $0.00035 per judgment, the interesting shift is coverage, not just savings.Online evals can move from sampling a few traces to scoring far more of them.Would you judge every production run at that price?
Open-weight models handled 56% of Vercel AI Gateway tokens in August, but only 14% of spend.
That split matters more than another “best model” ranking.
Where are you already using open-weight models in production?
Spend share can move even when routing behavior barely changes. Price and token mix distort the signal.
The more useful gateway metric is request share by task type: are developers actually rerouting, or just spending more on a different mix?
Gemini 3.8 Live makes async function calling the default.
In production, the conversation can keep moving while tools are still running.
Treat turn state and tool state separately—or late results become hard to reason about.
Auth failure should terminate the request, not degrade into a valid session object.
LiteLLM fixed CVE-2026-59822 in 1.84.0, and CISA added it to KEV.
If your MCP endpoint is exposed: patch first, then review auth activity.
If your MCP endpoint is exposed: patch first, then review auth activity.
OpenAI's Agents API is now in public beta with durable sessions, compaction and recovery.
Agent infra is becoming a managed runtime. The design question shifts from keeping the loop alive to deciding which state your app still needs to own.
Google's ADK for Kotlin 1.0 adds resumable sessions, context compaction, and human-in-the-loop flows.
The interesting part isn't Kotlin parity: mobile agents have to survive process death.
Recovery should restore state safely, not just reload chat history.
Coding agents can accelerate experiments, but faster iteration raises the cost of weak provenance.
Model, prompt, tools, data snapshot, and code revision should travel with every run.
Otherwise faster research becomes harder-to-reproduce research.
Searchable agent history becomes part of the build input, not just documentation.
If search results aren't pinned to a repository revision and permission scope, the same codebase can inherit different context with no visible diff.
An agent needs to know which pattern your team settled on and why a change landed. That context no longer sits in just the code. Today we’re releasing Agentic Search: one API across every repo your agent can access, returning the code, the session, transcript, and prompt behind the change.
Built with @turbopuffer
entire.io/blog/introducing-a…
Gemini 3.8 Flash is now GA for long-horizon software engineering and autonomous agents.
GA is a release state, not a regression test.
Before swapping it into an agent loop, replay tool calls, long tasks, and recovery against your current baseline.
An autonomy ladder without rollback is incomplete.
When prompts, tools, routing, and permissions evolve together, the last verified workflow should be recoverable as one versioned unit.
30 features of an AI native company:
1) Function-by-function process blueprint of your entire business.
2) Everyone in org using a daily driver harness like Grok Bot, Claude Cowork, ChatGPT at Work.
3) Centralized intelligence layer that aggregates structured and unstructured data, documents, and business logic into a single source of truth that is queryable & agentic work can be done on top of.
4) Model routing via OpenRouter, Ramp, etc that optimizes cost-per-successful-task across the business.
5) Treat context as code, ensuring architecture documents and conventions remain updated while allowing for diligent upfront planning.
6) Willing to throw away everything that you've built every three months and reimagine all your workflows.
7) A “skills distribution system” is used to manage agent behavior and optimize for token efficiency by ensuring developers trigger consistent skills throughout their workflow.
8) Separate technical implementation from high-level specifications, enabling non-technical staff to contribute in a format that agents can utilize to build technical implementation plans.
9) A key software metric is “cost per accepted PR”, with a focus on driving these costs down through better token efficiency.
10) An automated, agent-native development system where fleets of AI coding agents handle planning, writing, testing, reviewing, and shipping code while humans define the intent and acceptance criteria.
11) Heavy planning with higher-effort models and executing with cheaper, faster models.
12) Agent harness that uses CLI tools to parse metadata within markdown files to traverse dependency relationships, allowing agents to be granular in their input token usage.
13) Finance org that runs processes continuously in accounting (record-keeping) to re-define/reset forecasts on a much, much tighter cadence.
14) Financial models embedded in the underlying OS across the org to help drive reasoning.
15) Citizen Developer SDLC where non-technical employees can take a solution from idea to production with governance, access, versioning, and software conventions built in.
16) Closed loop, self-improving non-engineering workflows that learn from previous runs based on external performance metrics or internal evals.
17) AI ROI framework that includes experimental phase, scaling phase, and optimizing phase with bets sitting in 3 buckets: infrastructure, innovation, and efficiency.
18) Paid marketing motion that uses agent swarms to deploy thousands of pieces of creative for testing, before increasing spend on human-generated ads.
19) AEO/SEO engine that audits, rewrites, and (ideally) generates SEO/AEO-optimized blogs on a weekly basis, and then measures if any of it worked.
20) Agentic cyber security solution that fights AI with AI.
21) Combo of RL gym and first-party data to fine-tune open source models on high-volume processes that need SOTA performance at reasonable cost.
22) Human touch and judgement gets reserved for the first and final mile of most processes.
23) Evals are core infrastructure of your business. Anytime new models come out you have an apparatus for testing cost & performance against core processes.
24) Everyone is a builder. Especially C-level execs.
25) Everything gets recorded because what you don’t capture can’t be turned into ai-enabled work.
26) Legal, HR, and IT work in lockstep with owners of AI agenda so that business’ ass is sufficiently covered without slowing down transformation.
27) Bias to disrupting yourself before being disrupted by others.
28) Guardrails before features. Agents inherit the permissions of whoever is asking, enforced in the data layer.
29) Earned autonomy. Feedback feeds the evals that gate each new version, and agents move up a ladder as they clear it: observe, suggest, act with approval, act alone. The endpoint is agents running whole workflows inside a defined boundary, with humans setting the standard instead of checking every answer.
30) Traceability as the training signal. Trace every output to its prompt, model, data, and approver, so human feedback attaches to something specific rather than a vague sense that something is off.
What's missing?
OpenAI plans to wind down model access through Cursor, with Nov 12 proposed as the cutoff.
The engineering lesson isn't the vendor drama: model access is a dependency.
Fallback paths should be tested before they're forced.
Generated architecture diagrams are useful only if they stay tied to the code.
If the graph isn't tied to a commit and regenerated in CI, reviewers can end up trusting a stale picture of the repository.
Should stale diagrams fail the build?
Stop manually making flowcharts.
Try this GitHub skill with Claude Code🤯
github.com/tt-a1i/archify