Gabriel Anhaia

Senior Software Engineer · Backend, architecture & AI agents · Author of several books about software engineering and AI

GitHub:

[Gabriel Anhaia @anhaiagabe]

https://t.co/NZAagf3t4o

[Gabriel Anhaia @anhaiagabe]

206 tokens vs 72: the same tool result, detailed vs concise, in Anthropic's example.

Tool output fills your agent's context. Offer a response_format option, return names over UUIDs, make errors say what to fix.

[Gabriel Anhaia @anhaiagabe]

https://t.co/sj70I6aHZ3

[Gabriel Anhaia @anhaiagabe]

AI can double your PRs. Not your reviewers.

One company's 2x mandate, 802 devs: merged PRs per dev hit 2.09x baseline.
Before: 89% of PRs got a human review.
After: 68%. Per-reviewer load roughly doubled.

[Gabriel Anhaia @anhaiagabe]

https://t.co/v0ZVq6gUwE

[Gabriel Anhaia @anhaiagabe]

Your first agent eval suite can be small. Anthropic's advice:
1. 20-50 simple tasks from real failures
2. Mostly grade what the agent produced, not its path
3. Read the transcripts

[Gabriel Anhaia @anhaiagabe]

https://t.co/TGqkM4xDhA

[Gabriel Anhaia @anhaiagabe]

Most MCP clients load every tool definition before your agent starts.

Before: all tools upfront, all results via the model.
After: tools as files, read on demand. Anthropic's example: 150,000 tokens to 2,000.

[Gabriel Anhaia @anhaiagabe]

https://t.co/oiJriLHzk9

[Gabriel Anhaia @anhaiagabe]

Keep coding-agent tasks small if you want the PR merged.

A study of 33,596 agent PRs on GitHub: 71.48% merged. The unmerged ones tended to change more files and lines, and to fail more CI checks.

[Gabriel Anhaia @anhaiagabe]

https://t.co/6ofUtrfgjZ

[Gabriel Anhaia @anhaiagabe]

OpenAI had 93 developers screen 1,699 SWE-bench tasks. 68.3% were cut.

Flagged: 61.1% for tests that may reject a valid fix, 38.3% for underspecified issues.

Read your eval's tasks before trusting its score.

[Gabriel Anhaia @anhaiagabe]

https://t.co/WHNk1Ciixn

[Gabriel Anhaia @anhaiagabe]

Evals and backfills can wait. Batch them for 50% off on Anthropic's API.

1. Results arrive in any order. Match on custom_id
2. Unfinished after 24h? Expired, not billed
3. Cache hits: best-effort

[Gabriel Anhaia @anhaiagabe]

https://t.co/bcKHe5ydkE

[Gabriel Anhaia @anhaiagabe]

On a reasoning model, max_output_tokens caps the hidden thinking too.

Run out mid-thought: you pay for the reasoning and get "incomplete", no text.

OpenAI: reserve 25,000+ to start.

[Gabriel Anhaia @anhaiagabe]

https://t.co/rAxbOLSkMU

[Gabriel Anhaia @anhaiagabe]

Experienced devs don't vibe with coding agents. They steer.

13 observed, 99 surveyed, all 3+ years pro:
1. They plan before the agent codes
2. Median 1.8 plan steps per prompt
3. 9 of 13 observed reviewed every change

[Gabriel Anhaia @anhaiagabe]

The shape is guaranteed. The values aren't.

OpenAI's own docs: outputs can still contain mistakes, and a refusal may not follow your schema.

Validate values in code. Handle refusals.
https://t.co/2rLsTSysoP

[Gabriel Anhaia @anhaiagabe]

Valid JSON is not the same as your schema.

JSON mode gives you valid JSON. Structured Outputs also enforces your schema: no missing required keys, no invalid enum values.

[Gabriel Anhaia @anhaiagabe]

Constrained decoding: the schema becomes a grammar, and each step only allows tokens that keep the output valid.

The Outlines paper (Willard & Louf, 2023) does it with a state machine and a vocabulary index, at little added overhead.