TECHA working developer's toolkit for Claude CodeTRENDSThe quiet convergence: when agent frameworks start looking alikeTECHRAG is dead, long live RAG — why retrieval isn't going anywhereINFOWhy Claude's extended thinking changes the economics of complex tasksTECHA field guide to 2026's LLM benchmarks (and the ones you should ignore)TRENDSFine-tuning vs. long context: a practical decision matrixTRENDSWhat Anthropic's memory API tells us about the next year of agentsTECHPrompt caching is the closest thing to free compute we haveTECHA working developer's toolkit for Claude CodeTRENDSThe quiet convergence: when agent frameworks start looking alikeTECHRAG is dead, long live RAG — why retrieval isn't going anywhereINFOWhy Claude's extended thinking changes the economics of complex tasksTECHA field guide to 2026's LLM benchmarks (and the ones you should ignore)TRENDSFine-tuning vs. long context: a practical decision matrixTRENDSWhat Anthropic's memory API tells us about the next year of agentsTECHPrompt caching is the closest thing to free compute we have
AI × Forward
Subscribe
Trends

The quiet convergence: when agent frameworks start looking alike

After eighteen months of wild variance, the top five agent frameworks are converging on the same four primitives. What that means for the next wave of tools.

Initial Editor·2026-04-18·3min read·246 words

The shape everyone's landing on

Spend a weekend reading the changelogs of the top five agent frameworks and a pattern emerges. Strip the marketing away and each one now exposes roughly the same four primitives:

  1. A typed tool interface — JSON schema in, JSON out, with optional streaming.
  2. A scratchpad / memory store — opaque to the model, queryable to the runtime.
  3. A step loop with configurable stopping criteria (budget, tool calls, time).
  4. A trace / event stream that a host app can subscribe to for UI and debugging.

None of this is surprising in isolation. What's interesting is the convergence: these four were contested a year ago and they aren't anymore.

What this unlocks

Once the shape is shared, two things happen. First, tool authors stop shipping one adapter per framework. Second, tracing and eval tooling becomes portable — a win that's hard to overstate if you've been maintaining five similar integrations.

The next frontier isn't another primitive. It's defaults — who ships the best opinion about retries, about cost caps, about when to escalate to a bigger model. Expect that fight to define 2026.

What to watch

  • How each framework handles structured outputs inside tool calls. That's the only remaining shape mismatch.
  • Whether OpenAI's Assistants API stays in its own lane or finally adopts the same primitives.
  • Whether the "agent" word survives another year; several teams are quietly renaming this category to just "loops."

// more in trends

see all →